A comprehensive survey on human pose estimation approaches

Abstract

The human pose estimation is a significant issue that has been taken into consideration in the computer vision network for recent decades. It is a vital advance toward understanding individuals in videos and still images. In simple terms, a human pose estimation model takes in an image or video and estimates the position of a person’s skeletal joints in either 2D or 3D space. Several studies on human posture estimation can be found in the literature, however, they center around a specific class; for instance, model-based methodologies or human movement investigation, and so on. Later, various Deep Learning (DL) algorithms came into existence to overcome the difficulties which were there in the earlier approaches. In this study, an exhaustive review of human pose estimation (HPE), including milestone work and recent advancements is carried out. This survey discusses the different two-dimensional (2D) and three-dimensional human (3D) pose estimation techniques along with their classical and deep learning approaches which provide the solution to the various computer vision problems. Moreover, the paper also considers the different deep learning models used in pose estimation, and the analysis of 2D and 3D datasets is done. Some of the evaluation metrics used for estimating human poses are also discussed here. By knowing the direction of the individuals, HPE opens a road for a few real-life applications some of which are talked about in this study.

Full text

Vol.:(0123456789)
1 3
Multimedia Systems
https://doi.org/10.1007/s00530-022-00980-0
REGULAR PAPER
A comprehensive survey onhuman pose estimation approaches
ShradhaDubey1· ManishDixit1
Received: 23 November 2021 / Accepted: 2 July 2022
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2022
Abstract
The human pose estimation is a significant issue that has been taken into consideration in the computer vision network for
recent decades. It is a vital advance toward understanding individuals in videos and still images. In simple terms, a human
pose estimation model takes in an image or video and estimates the position of a person’s skeletal joints in either 2D or 3D
space. Several studies on human posture estimation can be found in the literature, however, they center around a specific
class; for instance, model-based methodologies or human movement investigation, and so on. Later, various Deep Learning
(DL) algorithms came into existence to overcome the difficulties which were there in the earlier approaches. In this study,
an exhaustive review of human pose estimation (HPE), including milestone work and recent advancements is carried out.
This survey discusses the different two-dimensional (2D) and three-dimensional human (3D) pose estimation techniques
along with their classical and deep learning approaches which provide the solution to the various computer vision problems.
Moreover, the paper also considers the different deep learning models used in pose estimation, and the analysis of 2D and 3D
datasets is done. Some of the evaluation metrics used for estimating human poses are also discussed here. By knowing the
direction of the individuals, HPE opens a road for a few real-life applications some of which are talked about in this study.
Keywords Human pose estimation· Deep learning· 2D/3D pose estimation· Activity recognition
1 Introduction
Human pose estimation (HPE) is described in the images or
videos as the problem of locating human joints (keypoints
such as elbows, wrists, etc.). It is one of the challenging tasks
in the area of computer vision, its interpersonal occlusions,
perfect estimation of the joints, and clothing make it difficult
to estimate. In addition to this, it has several applications in
real-life scenarios. Basically, it is a method of estimating the
body configuration (pose) from a single, usually monocular
image [1]. The problem of HPE may depend on many axes
such as single person or multi- person human pose estima-
tion, as the name suggests this estimation depends on the
number of people present in a frame. The other factor on
which HPE counts can be human body models which rely
on shape, cone, and mesh-based models. The way in which
the features are extracted also plays a vital role in estimat-
ing the human poses. The pose estimation for deep learn-
ing consists of views which are a holistic view (depends on
joint position regression), local view (body parts detection),
combining global and local information (body parts detec-
tion + joint position regression), and many others like motion
features, pose recognition in videos, foreshortening [2]. In
recent times, while digging deeper into this area, research-
ers are facing numerous obstacles like disorganization of
background, complication in poses, foreground occlusion,
and lack of robustness. Furthermore, high space and time
complexity also restrict its application value. Therefore,
various deep learning algorithms have been introduced to
overcome these challenges. This paper extensively reviews
how the HPE is enhanced from traditional ways to deep
learning-based 2D/3D human posture estimation strategies
from images or video recordings.
Communicated by R. Huang.
* Shradha Dubey
dubeyshradha29@gmail.com
Manish Dixit
dixitmits@mitsgwalior.in
1 Department ofComputer Science andEngineering, Madhav
Institute ofTechnology & Science, Gwalior, MP, India

S.Dubey, M.Dixit
1 3
1.1 Problem definition
This is a difficult task due to strong articulations, small and
hardly visible joints, occlusions, clothes, and illumination
variations.
There are many more different problem axes on which
human pose estimation models can be classified-
• Type of input modality (Red–Green–Blue (RGB)
image, Depth (Time of Flight) image, Infra-red (IR)
image)
• Number of cameras (Singleview or multiview)
• Human body models (Kinematic, planar, or volumetric)
• Type of images (Static or frames from video sequences)
• Number of people being tracked i.e. (single person or
multiperson pose estimation)
• 2D/3D pose estimation
Each of these problem definitions is discussed in further
sections of the paper.
1.2 Motivation
• A complete evaluation of modern deep learning and
traditional 2D and 3D HPE approaches is offered by
organizing them based on a 2D or 3D environment.
• A brief overview of major deep learning models such as
openpose, deepcut, alphapose, mask RCNN, and Itera-
tive Error Feedback are given
• An examination of a variety of applications of 2D/3D
HPE, including gaming, training robots, medical, activ-
ity recognition, augmented/virtual reality, and anima-
tion.
• In consideration of key issues in HPE, an informative
assessment of 2D and 3D HPE is offered, with future
perspectives towards enhancing performance.
• In addition to this, different types of datasets that are
used for the detection of 2D/3D pose estimation and
performance metrics that help in evaluating a model are
discussed in detail. Moreover, a separate description of
2D and 3D datasets are mostly missing in many of the
surveys.
• The paper also reviews the different deep learning
adversarial security attacks which may harm the well-
implemented deep learning models.
The purpose of writing this survey is that it covers
almost all the aspects of HPE along with the applications,
challenges, research gaps, datasets, evaluation metrics, and
security attacks which makes it different from the other
studies in the literature. Thus, this paper is much more
different and better than the other studies as nowhere in
the published surveys are the discussion about the security
attacks and challenges. The main pains of this area i.e., the
effect of occlusion and crowd on time complexity, accu-
racy, and what are the measures to prevent this in an image
are separately taken into consideration. Thus, covering all
the areas in a single survey is not yet seen in any of the
surveys in the literature hence, it is important. Last but
not least most of the open areas for future researchers are
well defined so that one could easily find a way for future
findings in this particular field.
PE has comprehensive and multiple uses, response to
circumstances of body reality. Some of the applications of
HPE are as follows.
1.2.1 Activity recognition
Activity Recognition is an estimation of a person’s identity,
personality, knowledge in a range of application areas rang-
ing from user interfaces to health monitoring, security, and
protection. They have shown great advances, particularly for
the task of pose estimation, as they can draw an appropriate
feature when discriminating together [3]. One of the key top-
ics for studying the science of computer vision and machine
learning is the human capacity to perceive the actions of
another.
Chengjun Chen etal. [4], The current study looks at how
object recognition and pose prediction technology can be
used in computer vision to identify a worker's assembly
action. Furthermore, an efficient deep learning algorithm is
used to estimate the running times of repeatable and tool-
dependent assembly behavior. Instead of using the traditional
behavior recognition algorithm, the YOLOv3 object detec-
tion algorithm is used in this analysis. It was discovered that
the YOLOv3 scheme has a major impact on assembly action
identification, with an accuracy rate of 92.8%. In addition,
the CPM pose estimation algorithm is used to obtain the
human body's joint knowledge.The judgment of running
times for repeated assembly actions is 82.1% accurate.
Murilo Varges etal. [5], suggested a new way of depict-
ing 2D poses in our paper. The 2D position is first translated
to parameter space, where every other segment is allocated
to a point, rather than utilizing straight line segments sim-
ply. Then, using a Bag-of-Poses technique, spatiotemporal
characteristics from the parameter space are retrieved and
encoded, and utilized to recognize human action in the
movie. Experimental studies on two well-known public
datasets, Weizmann and KTH, revealed that using 2D poses
encoded in the spatial domain, the proposed approach would
increase recognition rates while maintaining competitive
accuracy rates as compared to other methods.
Francisco Javier Ordonez and Daniel Roggen [6], This
paper are centered on convolutional and LSTM recurrent

A comprehensive survey onhuman pose estimation approaches
1 3
units, and a generalized deep architecture for behavior rec-
ognition is suggested. This network is appropriate for mul-
timodal wearable sensors; automatically performs sensor
fusion; does not necessitate professional expertise in feature
design, and specifically models the temporal dynamics of
feature activations. The system is tested on two databases,
one of which was used in a public activity identification
competition. The framework works better than other deep
learning models defined for activity recognition. Using
only accelerometers, the model achieves a 0.69 F1 score.
As accelerometers and gyroscopes are combined, accuracy
increases by 15% on average, and by 20% when acceler-
ometers, gyroscopes, and magnetic sensors are combined.
This provides a trade-off for applications that include the
automated processing of sensor data from various sources.
1.2.2 Medical
The pose measure has been used to classify postural prob-
lems, such as scoliosis, through the examination of irregu-
larities in the posture of the patient and physical therapy [7].
Kenny Chen etal. [8], suggested a semi-automated
method for optimizing upper-body pose prediction in noisy
clinical areas, in which the expansion and development
of a jointmonitoring system (Caffe Heatmap) are done to
enhance its reliability to uncertainties. The developed system
employs subject-specific CNN models trained on a segment
of a sufferer's RGB video recording selected to optimize
each joint's function variance. Besides that, the expanded
system yields more robust and precise posture annotations
by remunerating cene lighting variations and optimizing the
expected joint trajectories through a Kalman filter with fitted
noise parameters. The system achieves an average accuracy
of 96.5%, 96.8%, and 82.5% in all three subjects.
1.2.3 Training robots
Care robots are increasingly anticipated in the light of rising
birth rates, an aging population, and a shortage of care staff.
For instance, robots are supposed to monitor the condition
of residents when patrolling the facility while taking care of
nursing homes and other such establishments. Although the
initial estimate of a position (standing, sitting, falling, etc.)
is useful for assessing the condition of an individual, most
methods to date use photos.
Md Jahidul Islam etal. [9], a framework presented for
determining the 3D relative pose of pairs of talking robots
using human pose-based key points as correspondences.
To begin, an OpenPose is employed to extract pose-based
2D key points for the humans in the image. Following that,
an iterative optimization algorithm is used to optimize the
key-point correspondences based on their local structural
similarity in the image space. Finally, the follower robots
solve the perspective-n-point (PnP) problem to align the
corresponding 2D predictions on their respective measured
cameras and locate their relative poses. The model achieves
an accuracy of 75.67, 57.82, and MAP 72.26, 54.91 on the
Market-1501 dataset CUHK-03 dataset respectively.
Zimmermann etal. [10], developed a CNN-based method
for predicting 3D human pose in real-world units by combin-
ing color and depth detail. The outperforms monocular 3D
pose estimation approaches based on color and pose estima-
tion based solely on depth. The method works on two data-
sets namely, MKV and captures dataset. The device is used
in conjunction with a demonstration-based learning platform
to teach a service robot without the use of markers. Studies
in real-world environments show that the proposed method
allows a PR2 robot to mimic the manipulation behavior of
a human instructor.
Vasileiadis etal. [11], This paper explores a human pose
prediction and monitoring system capable of precise and
stable real-time output in real-life applications, with a focus
on robotic-assisted living application fields. The method
extends the effectiveness of the articulated-SDF monitor-
ing approach by adding a set of complementary frameworks
to address issues that occur in real-world situations, such as
tracker activation and malfunction, pose monitoring from
partial-views, broad object management, and body compo-
nent convergence. Three publicly accessible human pose
tracking datasets, SMMC-10, EVAL, and PDT, are used to
test the overall evolved pose tracking process. On a practi-
cal human motion dataset, the experimental system achieves
positive results: ADE 0.075, mAP 0.825, and outperforms
the Kinect built-in pose estimator.
1.2.4 Animation
Character animation was historically a manual process.
Nevertheless, postures may be specifically matched with
an individual actor by advanced posture evaluation frame-
works. More established frameworks depended on markers
or special suits. Late advances in pose estimation and motion
selection have made unmarkable, often real-time applica-
tions [12].
Kumarapu and Mukherjee etal. [13], In this work, Ani-
mePose is introduced, a supervised multi-person 3D poses
estimation and animation system for a given RGB video set.
The work is divided into various modules like first estimat-
ing 2D poses then lifting 2D poses 3D poses. The work is
done on a publicly accessible MuPoTS-3D dataset, the pro-
posed solution achieves similar outcomes to previous state-
of-the-art 3D multi-person pose prediction approaches, and
it also outperforms previous competitive human pose moni-
toring methods by a substantial margin of 11.7 percent accu-
racy benefit on MOTA score on the Posetrack 2018 dataset.

S.Dubey, M.Dixit
1 3
Mohit Tiwari etal. [14], a method is introduced for esti-
mating a human’s 2D pose from a video that is both fast
and effective. The method employs an affinity vector field
(AVF) that discovers the relationships between body parts.
The architecture uses global representation codes, allowing
for a bottom-up approach with high accuracy in real-time.
Moreover, the architecture is designed to understand and
associate part locations together. The MPII dataset is used
in this study.
For the application of Human Pose, Evaluation Amazon
Go presents an important area. The camera track recognizes
people and their actions, which is an important component
of Pose Estimation. The services that monitor and measure
human activities are largely based on an estimate of human
employment. Other applications of HPE are gaming, motion
tracking Casado García etal. [15], motion capture, aug-
mented reality, and many more.
Albert Cleetus [16], In this paper, deep CNN is used for
developing motion capture for animation. The images are
captured live which will be helpful in real-time pose esti-
mation for gaming applications. The captured video is then
passed through various steps before mapping to 3D character
animation. These steps are object detection, cropping region
of interest, 3D pose reconstruction, and data normalization.
The developed model can be used with ease and is afford-
able. The methods were put to the test on 10,000 images,
and the results of the 3D pose prediction were compared to
the GT data.
1.3 Challenges
The main challenges for human pose estimation area vari-
ation of body poses, complicated background, and depth
ambiguities. To solve these problems, considerable research
efforts have been devoted to the related fields.
1.3.1 Lack ofaccuracy
The ultimate goal of every type of model is to achieve a
better accuracy i.e., must perform its task efficiently and
precisely. Therefore, the implemented algorithm must be
good at its accuracy. Most of the time, it has been observed
that accuracy has been neglected as one is focusing on other
parameters while implementing an algorithm one may
neglect accuracy.
1.3.2 Occlusion
Occlusion can occur in both single and multiple human pose
estimation. It mainly creates problems in detecting accurate
human poses in an image or video because the part of the
human body gets occluded thus becoming an obstacle while
detection. Thus, different occlusion handling techniques are
proposed to overcome this problem.
1.3.3 Crowd
Alike occlusion, crowd also becomes an obstacle in detect-
ing poses but it can occur only in the multiperson human
pose estimation approach.
1.3.4 Time complexity
The algorithm should be designed in such a way that it
should be time-efficient. It has been observed from the
literature that most of the traditional approaches require
much time to achieve better results and some of the deep
learning approaches also do not have good results in case
of time complexity if the selected data is not proper. Thus,
time complexity constraints must always be in mind while
designing an algorithm.
1.3.5 Preprocessing
Preprocessing is also one of the most challenging parts of
HPE. The localization of body parts, background subtrac-
tion, data calibration, and image conversion thus plays a
major role while detecting poses.
1.3.6 Data security
One of the most common concerns in ML/DL is security and
privacy. Once a corporation has uncovered the data, security
is a critical concern that must be addressed. To execute this
accurately and efficiently, distinguishing between sensitive
and insensitive data is critical.
These are some of the challenges which should not be
ignored while implementing a human pose estimation
algorithm.
1.4 Paper organization
The paper is divided into many sections. Section2 gives a
brief about the different human body models as they play a
fundamental role in detecting different human poses. Next,
Sect.3 discusses the classical and deep learning approaches
of 2D pose estimation and its classification into a single
person and multiperson pose estimation. The single-person
pose estimation is further classified into direct and heat-
map regression. In the summary of 2D pose estimation,
the comparison of all these approaches is done along with
their advantages and disadvantages. Similarly, in Sect. 4,
classical ad deep learning approaches to 3D human pose
estimation are discussed and these approaches are also
divided into monocular, a single person, and multiperson

A comprehensive survey onhuman pose estimation approaches
1 3
3D pose estimation, and at last, the comparison of all these
approaches is done.
In Sect.5, some of the major deep learning models for
HPE are discussed. Further in Sect.6, the separate descrip-
tion of each of the 2D and 3D datasets are given, and also
the evaluation metrics are discussed, at last, the paper binds
up with the conclusion and future work for further research
on this particular topic.
2 Human body models
The position of human body parts is required to develop a
human body representation from visual input data in human
pose estimation [17]. As a result, human body models are a
critical component of human pose estimation. It addresses
highlights and key points extracted from input images. To
define and predict human body postures and generate 2D
or 3D poses, a model-based technique is typically applied.
There are three types of pose estimation based methods
(Fig.1).
• Kinematic modeling: This is also known as a skeleton-
based model, and it is used to estimate 2D and 3D poses.
To reflect human body structure, this adaptable and
comprehensible human body model comprises a vari-
ety of joint configurations and limb orientations. As a
result, this model is utilized to represent the relationships
between various body parts [18]. The kinematic model,
on the other hand, has limitations when it comes to cap-
turing texture or shape information.
• Planar model: It’s also referred to as a contour-based
model, and it’s primarily used to represent 2D body con-
tours. This model is used to illustrate the human body’s
appearance and shape. Usually, body parts are repre-
sented by multiple rectangles approximating the human
body contours [17].
• Volumetric model: Typically, body parts are depicted by
a series of rectangles that resemble the contours of the
human body.
Figure2 shows all three types of human body models. a
shows how the various joints are determined with the help
of key points represented by solid circles thus, helping in
estimating poses. Similarly, b shows the rectangles.
3 2D pose estimation
The placement of key points in 2D space relative to an image
or video frame is readily estimated with 2D pose estimation.
It works by detecting and analyzing the X, and Y coordinates
of human body joints in a picture [3]. 2D Pose Estimation is
the process of identifying the location of body joints in an
image (in terms of pixel values).
3.1 Classical approaches
There are different ways to deal with 2D human pose estima-
tion and some general use classical approaches, for exam-
ple, HOG, Edgelet, pictorial structure model (PSM) [19–21].
Some of the traditional works in the literature for analyzing
the human stances are evaluated by making use of figure
drawing for example by drawing cylinders for each body part
and assessing the posture by joining these cylinders.
The pictorial structure parts are defined by pixel area and
direction, which brings about the prevailing methodology for
human pose estimation [19]. Pictorial structures have been
proposed by M.A. Fischler [20]. It is a rearranged approach
to portraying an item. The model has two components that
Fig. 1 Classification of Human body models
Fig. 2 Human body models
from the left skeleton-based
model; contour-based model;
volume-based model [1]

S.Dubey, M.Dixit
1 3
comprise 2D image sections and structure which is an
assortment of parts. An effective answer for pictorial struc-
tures was proposed by [2] as shown in Fig.3. They showed
how dynamic programming can be efficiently processed with
pictorial structures if the representation has no cycles. It
demonstrates 2D face recognition and human pose estima-
tion applications. Thus, pictorial structures are appropriate
for 2D human posture estimation. The above method, on
the other hand, has the drawback of requiring a pose model
that is independent of image data. Consequently, research
has concentrated on improving the models’ representational
capability.
Interestingly, conventional models for object recognition
use parts defined exclusively by areas, which disentangles
both deduction and learning. Such models have been demon-
strated to be exceptionally fruitful for object location. Along
these lines, Yang [21] proposes a blended model of parts that
presents complex connections, having a unary layout for and
use in the assignment of recognizing explained individuals,
and assessing their postures. The Part-based model excels at
simulating articulations. However, this comes at the expense
of restricted expressiveness and does not account for the
global context (Fig.4).
There are many shortcomings in traditional strategies,
such as the tree model does not reflect all the limitations,
the extraction execution of handcraft highlights is low and
the complexity of reasoning is high. Due to certain param-
eters, the accuracy of the model gets disturbed and thus the
model is not able to perform according to the expectations.
Because of the limitations of conventional methodologies
and to improve the presentation of assessing the postures,
deep learning models have emerged, different models are
dependent on the single person or multiperson. Some of the
work based on deep learning approaches are reviewed below
(Fig.5).
3.2 Deep learning‑based approaches
The traditional pipeline contains flaws, and pose estimation
has been drastically altered by deep learning-based method-
ologies. These techniques are categorized as single person
and multi-person pose estimation methods.
Fig. 3 Pictorial structural model
[20]
Fig. 4 Examples of a HOG features for keypoint detection [21] and contour Features from [22] b An original image; c Extracted contours

A comprehensive survey onhuman pose estimation approaches
1 3
3.2.1 Single‑person pose estimation (SPPE) methods
Single-person techniques detect a specific person’s pose in
an image. When an image contains multiple people, then the
image is cropped in such a way that there is only one person
left in the image. An upper-body detector [23] or a full-body
detector [24] can perform this task automatically. The goal
of single-person techniques is to locate the keypoint loca-
tion in that area based on the given position information.
Depending on how they predict key points, an SPPE pipeline
is categorized into two categories.: KeyPoint Regression-
based approaches and Heatmap-based approaches.
3.2.1.1 KeyPoint regression The model regresses the key
body points directly from the feature maps in this manner,
which is referred to as Direct Regression in some literature.
The model's output will be a 17 by 2 vector comprising the
X and Y coordinates of each anticipated key point if you
want to estimate 17 key points for an individual using this
method as shown in Fig.6.
Several various models are presented to boost the effec-
tiveness of the keypoint regression strategy in getting the
correct points, such as Jaio Carreira etal. [26] suggested the
Iterative Feedback method, which uses feedback to develop
predictors that can efficiently manage complicated, struc-
tured output spaces. The primary goal is to learn several
levels of feature extractors throughout their joint space to
derive the 2D regions of a variety of key points from an
RGB image, such as the ankle joint, shoulder, and so forth.
Rogez etal. [27] developed the LCR network, in which each
individual is first located, then categorized using a set of
anchor postures, and finally regressed. A downside of this
method is the large number of posture anchors required to
achieve reliable results. Toshev and Szegedy [28] proposed
and introduced a cascaded DNN regressor for directly pre-
dicting human keypoints. However, understanding mapping
Fig. 5 The taxonomy used in this section
Fig. 6 KeyPoint regression [25]

S.Dubey, M.Dixit
1 3
straight from feature maps without the use of other processes
is difficult.
3.2.1.2 Challenge Imagine a situation where the model
chooses a specific key-point location with a deviation of
one or two pixels from the underlying data. This little dis-
crepancy in the model’s estimation generates an error that
disrupts the training process and prohibits the model from
settling into an optimal solution; nonetheless, in many appli-
cations, even a small divergence in estimation is tolerable.
As a result, training a model to explicitly identify the exact
location increases the problem’s complexity and sensitivity,
as well as the model’s training instability.
3.2.2 Heat map regression
Heatmap-based frameworks are now widely used in 2D
HPE tasks. Heatmap-based framework regresses heatmaps
first. In this framework, the probability of the existence of
a key point in each pixel of the image is estimated. A more
probable keypoint zones using a heat map is demonstrated.
Many researchers have used heatmap regression models
like Shih-En Wei etal. [29] propose Convolutional Pose
Machines (CPMs) for the task of articulated pose estima-
tion that consistently produces 2D conviction maps. At
each phase in a CPM, features of the image and the con-
viction maps created by the past stage are given as input.
The proposed multi-stage design is completely differenti-
able and accordingly can be prepared by utilizing back-
propagation. The method is evaluated on the FLIC Data-
set [30] and achieves PCK@0.2 with 97.59% on elbows
and 95.03% on wrists. Newell etal. [31] reviewed that
due to a stacked block architecture with numerous inter-
mediary supervisions, a stacked hourglass network with
convolutions in multi-level features was developed thus,
allowing re-evaluation of prior estimations. Adrian Bulat
etal. [32] suggested that even when there is considerable
partial occlusion, the human pose can be detected by fol-
lowing their approach of pose estimation detection. On the
heatmaps, this method employs detection and regression.
Recently, Zhengxiong Luo etal. [33] proposes the scale-
adaptive heatmap regression (SAHR) method, which can
modify the standard deviation for each keypoint flexibility.
SAHR is followed by weight-adaptive heatmap regression
(WAHR), which helps in balancing the foreground and
background samples. Extensive testing has shown that
combining SAHR and WAHR increases the accuracy of
bottom-up human posture assessment significantly. Chen
and Yuille [34] used DCNN to learn pairwise relations
using the graphical model (parts type and pairwise parts
relationships) (Fig.7).
3.2.2.1 Challenges Regression of heat maps faces two
obstacles. First, there’s the keypoint extraction issue utiliz-
ing heat maps (decoding problem), which can be solved by
selecting the maximum or average of each heat map as a
key-point position. The other issue is generating a ground
truth [36]; because the model outcome is in the form of a
heatmap, therefore, it is needed to convert the ground truth
(which is made up of keypoint coordinates) to the same for-
mat (encoding problem).
3.2.3 Comparison ofdirect regression vs heat map
regression
Direct regression of joint locations is very nonlinear, making
learning mapping challenging [27]. It also can’t be used in
a multi-person situation. The heatmap-based system, on the
other hand, regresses heatmaps first. Heatmaps can be used
to aid human comprehension and model more complex sce-
narios. Direct regression is more trustworthy and has certain
virtues because it is fast and simple [26–28]. When direct
regression is used, the ultimate result can be determined
from stem to stern without the need for heatmaps. It can also
be used in 3D scenarios with minimal adjustments. Further-
more, integrating heatmaps with huge convolutional kernels
Fig. 7 Pose estimation using heat map regression[35]

A comprehensive survey onhuman pose estimation approaches
1 3
and deeper models increases performance by expanding the
relevant receptive field and hence the context acquired. [29,
31]. As the training proceeds, the erroneous response map
is gradually suppressed, and the correct response map has
gradually become strong. As a result, there seems to be no
absolute answer to this question, and each paradigm has its
own set of benefits and drawbacks.
3.2.4 Multi‑person pose estimation
Because the position and volume of people in an image are
uncertain, multi-person pose estimation is more complex
than single-person posture estimation. In most cases, the
problem can be addressed in one of two ways:
• The simplest method is to start with human detection,
then estimate the parts, and finally calculate the pose for
every individual [37, 38]. This approach is called the
top-down approach.
• Another method is to identify all portions in the image
(i.e., parts of each individual), then group parts that
belong to different people. This is referred to as the bot-
tom-up approach [39].
3.2.4.1 Top‑down approach The top-down technique of
HPE works in stages: initially, it detects the person from
an observable frame, then it obtains that discovered object,
and finally, it attempts to estimate 18 critical points for
every identified person from in that frame before attempt-
ing to form the skeleton. He etal. merged segmentation and
keypoint prediction in the Mask-RCNN model. To obtain
a one-hot mask for each keypoint, the authors attach key-
point heads on top of RoI-aligned feature maps. For multi-
scale inference, Chen etal. built globalnet on top of Feature
Pyramid Networks and refined the predictions with hyper-
features. [38]. Papandreou etal. proposed a model which
works in two stages. Firstly, a faster RCNN detector is used
to detect the location and scale of boxes and then the key
points of the person are estimated [37]. In this approach, a
novel aggregation is used which is used to aggregate these
results and produce highly localized keypoint predictions
by combining these outputs. Nelson Rodrigues etal., make
use of time of flight (ToF) images and take out the ROI.
Detecting joints on padded ROIs did significantly change
the results and enabled the system to be more effective for
joints near the ROI boundary [40].
3.2.4.2 Bottom‑up approach On the other hand, the bot-
tom-up approach firstly tries to estimate the 18 key points
from the image for each person [39]. After that, it will try to
estimate a person from those key points i.e. joining the key
points and making the skeleton. Muhammed Kocabas etal.
[41] use a pose residual network (PRN) that takes keypoint
and the human detection process and assigns keypoints to
human instances to generate valid poses. Iqbal etal. offer
a new method for modeling multi-person pose computa-
tion and tracking in one formulation. The study resolves
the densely connected graphical model locally, resulting in
a significant increase in time efficiency. Insafutdinov etal.
proposed DeeperCut [42] which improves DeepCut using
deeper ResNet and employs image-conditioned pairwise
terms to get better performance.
3.2.4.3 Comparison of top‑down and bottom‑up
approaches In general situation, the top-down approach
consumes time much more than the bottom-up, because the
top-down approach needs N-times to pose estimation by
person detector results. However, the bottom-up approach
misses the opportunity to zoom into the details of each
person’s instance [38, 39]. As a result, there is a discrep-
ancy in accuracy between these two approaches. Deep neu-
ral networks have been used to investigate both bottom-up
and top-down techniques in the latest days. However, as the
present scenarios are taken into account, it is impossible to
say which strategy is better than the other. For multi-person
pose estimate assessment, accuracy and speed are two main
parameters.
Accuracy: In terms of the bottom-up pipeline, since the
network cannot acquire reliable features from the frames,
the scale variability of people might create issues with HPE.
During the training phase, the average resolution of a single
person in a bottom-up pipeline is relatively low than in a
top-down pipeline using the same network and GPU stor-
age[43]. As a result, what truly limits the accuracy of the
bottom-up pipeline could be an equipment limitation.
Speed: Every human pose in the top-down pipeline is
evaluated individually, which takes linear time as the num-
ber of people increases. Every person’s pose in the top-down
pipeline is evaluated individually, which takes linear time
as the majority of individuals increases. As a result, in the
bottom-up pipeline, increased speeds could be possible[43].
Therefore, deep learning-based approaches have achieved
a breakthrough in HPE by improving performance signifi-
cantly. In the following, section the comparative analysis of
different deep learning-based 2D HPE methods concerning
a single person and multi-person scenarios is done (Table1).
4 3D pose estimation
Determining the articulated 3D joint (keypoints) regions of
a human body from an image or video is known as three-
dimensional (3D) HPE [52]. It calculates the 3D posture (x,
y, z) of an RGB image. The purpose of 3D HPE is to use a

S.Dubey, M.Dixit
1 3
Table 1 Comparative analysis of 2D human pose estimation techniques
Technique Single person Multiperson Advantage Disadvantage Performance
Real-time open pose [44] – Bottom-up approach The first real-time system for
detecting key points. Further-
more, it merges the body and
the foot into a single model,
increasing the accuracy and
reducing the inference time
Greedy multi-person parsing fails
in highly crowded images
Generates strong body pose parses
and maintains efficiency regard-
less of the number of persons
Deep cut subset partitioning and
labeling problem (SPLP)
– Bottom-up approach Empirical results on the four
datasets show improvement in
both single person and multi-
person dataset
The presence of several occlu-
sions in the background may
degrade the result
The proposed model infers the
number of persons, poses, spatial
closeness, and part-level occlu-
sions all at once
RMPE alpha pose [45] – Top-down approach It improves pose estimation in
the context of imprecise human
boundary boxes and repetitive
detections
it would be interesting if the
framework is trained together
with the human detector in an
end-to-end manner
Works better in terms of accuracy
and efficiency
Object localization using CNN
[46]
Heatmap – Estimate the joint offset location
within a small region of the
image
The model is sensitive to poor
lightning conditions, cluttered
background
Quicker and more computationally
efficient
Convolutional pose machines
[29]
– Multiperson Without the use of explicit
graphical model-style inference,
the model delivers increasingly
advanced estimates for part
positions
Gets confused when multiple
people are nearby
Achieve state-of-the-art results on
standard benchmarks includ-
ing the MPII, LSP, and FLIC
datasets
Stacked hourglass network [31] – Repeatedly use Top Down
and Bottom-up
Continuous performance across
a range of conditions, includ-
ing obstruction and multiple
individuals in proximity
Sensitive to camera character-
istics
Low computational cost
Deep high-resolution network
HRNet [47]
Heatmap – Reliable, high resolution The multiresolution representa-
tions are complex
Less complex, cost-effective
Cascaded pyramid network
(GlobalNet, Refinenet) [16]
– Top Down A separate network is used to
simply handle hard keypoints
Only capable of dealing with
head pose changes, occlusion,
and illustration
Low Computational cost
Real-time lightweight open-
pose[48]
– Bottom-up Suitable for real-time perfor-
mance on edge devices
High Complexity Heavily optimized network design
and post-processing code
Deeply learned compositional
model (DLCM)[49]
– Top Down and Bottom-up Use compositional patterns for
pose estimation
Multiple humans cannot be iden-
tified in the same frame
Lower complexities, compactly
encode orientations, scales, and
shapes of parts
Transformer (TF) pose [50] Improved regression – Removes different problems such
as quantization error and non-
differentiable post-processing
Does not able to produce better
results for occluded images
Simple and direct framework
MoDeep CNN [51] – FLIC-motion Incorporates both color and
motion features to detect human
poses
Moving objects in the backdrop
can have a significant impact on
the outcome
Even extremely modest temporal
signals can boost performance
with only a slight rise in com-
plexity

A comprehensive survey onhuman pose estimation approaches
1 3
picture of a person to determine the XYZ coordinates of a
particular set of keypoints on the human body. 3D keypoints
are visually observed as follows (Fig.8).
Following the extraction of joint positions, the activity
assessment model examines a person’s posture and exam-
ines the person’s real motion in a series of frames from a
video stream. Generally, it is harder to recuperate 3D poses
from 2D RGB images because of more ambiguities in its
estimation and it requires huge space as compared to 2D
images [47].
The selection of the dataset images in a 3D pose is also a
challenging task in its estimation. Moreover, the algorithm
has to be invariant to many other factors such as textures,
the skin color of the selected image, imperfections in an
image background, human occlusions, and many more. The
traditional and recent deep learning approaches for three-
dimensional pose estimation are reviewed below.
4.1 Classical approaches
Pictorial structure models (PSM) are the accepted norm for
2D human posture estimation. It proposes a multi-view pic-
torial structures model that expands an ongoing advance in
2D estimation. The 3D PSM is non-exclusive and relevant
to both single and multiple human posture estimations. It
is essentially a generative model for pose estimation. PSM
does not give many accurate results as they have the condi-
tions between the output factors, subsequently, to consider
these conditions SVM Hanguen Kim etal. [54], Ke Chen
etal. [55] techniques are acquainted for the structured SVM
utilization to learn the mapping from segmentation features
to joint locations.
3D human stance estimation utilizing HOG highlights
portrays the state of the object and is utilized to examine
3D human posture steadily. As the HOG highlights are reg-
istered over the whole picture, the HOG features measure-
ment is high along these lines as shown in Fig.3 and hence
principal component analysis (PCA) is used on each HOG
block [30]. Therefore, the 3D human pose can be assessed
by the linear regression of HOG features.
This approach took first place in the inaugural COCO
2016 keypoints challenge, outperforming the previous state-
of-the-art outcome on the MPII MultiPerson benchmark by a
wide margin, both in performance and efficiency. MPII data-
set result on the testing subset achieves a Map of 75.6% and
results on the COCO 2016 keypoint challenge is 61.8% [52].
4.2 Monocular 3D human pose estimation
Among the most essential and difficult topics in computer
vision is vision-based monocular HPE, which wants to
acquire human postures from input images or frames. Due to
depth uncertainties and occlusions, 3D HPE from monocular
photos is inadequate.
In both 2D and 3D contexts, the monocular camera is by
far the most extensively utilized detector for HPE. Research-
ers have been able to expand their findings to 3D HPE due
to recent advances in deep learning-based 2D HPE from
monocular photos and videos. Specifically, deep learning-
based 3D HPE methods are further separated into two main
groups: single-view 3D HPE and multi-view 3D HPE.
Hallquist and Zakhor [56] use single-view pose estimation
using mobile devices with a distance of 10m over the query
images. Fei etal. [57], calculate physical distances among
individuals from just a single RGB image or video acquired
by a camera observing a 3-D scene from a stable viewpoint.
3D keypoints are generally deduced using single-view
photos, irrespective of methodology (image → 2D → 3 D
or image → 3D). Conversely, multi-view imaging can be
employed, in which each frame is taken from multiple cam-
eras focusing on the target object from various perspectives.
The multi-view approach improves the perception of depth
and is useful when certain sections of the body are oblite-
rated in the image [58]. As a result, predicted values become
even more precise.
Fig. 8 Keypoints in three
dimensions and their definition
[53]

S.Dubey, M.Dixit
1 3
Usually, this technique necessitates camera synchro-
nization. Nevertheless, other researchers show that even
asynchronous and uncalibrated video streams from several
cameras may be used to predict 3D keypoints. Furthermore,
there are other methods for estimating 3D human pose:
• To train a model that can infer 3D joints directly from
the images given. For Eg, A multi-view system named
EpipolarPose is trained to predict the locations of 2D
and 3D keypoints simultaneously. The fascinating part
is that it only requires 2D keypoints for training instead
of ground truth 3D data [58]. Rather, it creates the 3D
ground truth using epipolar geometry for 2D projections
in a self-supervised manner. It’s useful since a scarcity
of elevated 3D pose annotations is a major issue when
developing 3DHPE algorithms.
• Identifying 2D keypoints and then translating them into
3D is the most popular strategy because 2D keypoint
projection is well-studied, and using a pre-trained frame-
work for 2D predictions improves the system’s accuracy
rate [59]. Furthermore, several existing models offer
reasonable accuracy and speed of inference in real-time
(for example, PostNet, HRNet, Mask R-CNN, Cascaded
Pyramid Network [16, 41, 47, 60]).
4.3 Single‑person 3D pose estimation
Most works for estimating human pose for a single person
use a single image/video. Despite the ambiguity in the depth
dimension, models trained on 3D ground truth (GT) show
pretty good performance for the case of a single person with-
out occlusions. The figure below shows the direct estimation,
2D to 3D lifting approaches, and volumetric model approach
(Fig.9).
4.4 Multiperson 3D pose estimation
The main challenge in multi-person 3D pose estimation is
occlusions. Due to a limitation of suitable datasets, pro-
gress on multi-person 3D pose estimation is inherently
limited[62]. In addition, unfortunately, there are almost no
annotated multi-person 3D pose datasets like the Human3.6
datasets. Most multi-person datasets either do not have good
Fig. 9 Block diagram of a direct estimation approaches and b volumetric model approach for SPPE [61]
Fig. 10 Block diagram of a Top-down approach and b Bottom-up approach for 3D human pose estimation

A comprehensive survey onhuman pose estimation approaches
1 3
Table 2 Comparative analysis of 3D human pose estimation techniques
Technique Single person multiperson Advantage Disadvantage Performance
Occlusion-robust pose-maps (ORPM)
[64]
Monocular RGB The proposed method works well even
under strong inter-person occlusions
and human interactions better than
previous approaches
Inaccurate predictions when joints of
the same type are nearby
Overall pose quality is improved
Deep convolutional neural network
[65]
Monocular images The network used has disentangled
the dependencies among different
body parts which makes it simpler to
detect human poses
Not feasible for the multi-view pose
estimation
The network achieves significant
improvement over baseline methods
Single-view-multi-angle consistency
(SVMAC) [66]
single view It is desirable if the image contains
only one human object
Not feasible When there is an occlu-
sion in the images
Faster computation; Easy to implement;
Location-maps-based model [67] Equirectangular Images Implements the enhanced location
maps-based model which takes both
distorted and disconnected images
into consideration
Not much robust to all the values of
location map variances
the model indicates better performance
with respect to accuracy and computa-
tion complexity
3D pictorial structures (3DPS) model
[68]
Single person/Multiperson Self and natural occlusions can be eas-
ily handled
Doest not work with real-time images The 3D PCP scores for single human
and multiple humans is 76 and 75.6
respectively
Efficient Pose [69] Single-person Practically used in-the-wild images
even when end-to-end training is not
feasible
Can take only a single image at a time The 3D PCP scores for single human
and multiple humans is 76 and 75.6
respectively
Deep Depth Pose (DDP) model [70] Single view/Multiview Deep Depth Model may generate a
3D posture by linearly combining
prototype poses
The model is much more complex Provides better accuracy on the dataset
used
Fully Convolutional Model [71] Video Based The model is practical in scenarios
where motion capture is challenging
Confused with occlusions and light-
ning conditions
It improves performance when labeled
data is scarce
fully connected neural network with
the SMPL [72]
Multiview Improves image generalizations Cannot restore posture distortion the joints’ average error of the proposed
method is the smallest

S.Dubey, M.Dixit
1 3
ground truth or are not realistic. Abdallah Benzine uses Pan-
daNet (Pose Estimation and Detection Anchor-based Net-
work) for estimating multiperson 3D poses [63] (Fig.10)
(Table2).
5 Major deep learning models inhuman
pose estimation
5.1 OpenPose
OpenPose is the first real-time bottom-up multi-person
system to detect human bodykey points (several 135 key
points) on single images. Researchers at Carnegie Mellon
University proposed it [44] (Fig.11).
• Initially, an image is processed through a CNN network
to retrieve the input’s feature maps. The first ten layers
of the VGG-19 network are used in the model.
• The part confidence maps (CM) and part affinity fields
(PAF) are generated from the feature map using a
multi-stage CNN pipeline.
• Each branch’s forecasts are refined over successive
stages. Bipartite graphs are created between pairs of
parts using part confidence maps (as shown in the
above image) [74].
• In multi-stage CNN, The PAF refine Lt from the
extracted features of the base network F in the first set
of phases.
• The next step uses the prior layers’ output PAF to refine
the confidence map recognition.
Lt
=𝜑
t(
F,L
t−1)
,∀2≤t≤TP
,
The greedy approach is then used to analyze the final S
(CM) and L (PAF).
5.1.1 Advantages
• OpenPose is significantly more precise since it’s designed
to run on GPUs. OpenPose is free for non-profit usage
and can be transferred below these terms.
• High accuracy without impacting implementation quality
5.1.2 Disadvantages
• The fundamental disadvantage of OpenPose is that its
low-resolution findings restrict the level of information
in keypoint predictions.
• It does not provide any statistics about the depth and is
built on DNN, which demands a powerful machine.
• Speed and precision are slightly hampered.
5.2 DeepCut
DeepCut is a bottom-up approach for estimating multi-per-
son human poses. The authors tackled the job at hand by
identifying the significant information:
• Make a list of D body part candidates. This integration
gives all probable body part locations for each person in
ST
P=𝜌
t(
F,L
T
P
)
,∀t=TP
,
St
=𝜌
t(
F,L
T
P,S
t−1)
,∀TP<t≤TP+Tc
,
Fig. 11 OpenPose architecture [44]

A comprehensive survey onhuman pose estimation approaches
1 3
the image. Choose a subset of body parts from the list of
candidates
• Each selected body part should be labeled with one of the
C body part classes. Theseclasses represent the various
types of parts, such as “arm,” “leg,” “torso,” and so on.
• Body parts that belong to the same person should be
separated [75].
5.3 AlphaPose
AlphaPose is an accurate top-down multi-person pose esti-
mator, which is the first open-source system. The model is
used to solve the problem of multi-person pose estimation
in the wild [73]. The AlphaPose framework consists of three
components:
• Symmetric spatial transformer network (SSTN).
• Parametric pose non-maximum-suppression (NMS)
• Pose-guided proposals generator (PGPG).
• First, it has the bounding box proposals by the human
detector. According to the paper,the authors used a VGG
based SSD512 detector for detecting humans.
• These bounding box proposals are then fed into the Sym-
metric STN (Spatial Transformer Network) + SPPE mod-
ule. This step generates the pose proposals.
• Now, there is also see a Parallel SPPE module in Fig.12.
This module is used during training to avoid the local
minimums [76].
• The detected poses may also contain many redundant
detections. To reduce these, the authors used a parametric
Pose-NMSto eliminate the redundant poses.
• Also, to augment the training samples, the authors used a
Pose Guided Proposals Generator (PGPG)during train-
ing.
5.3.1 Advantages
• Designed to eliminate flaws in the traditional system such
as erroneous identification or localization.
• To boost performance, optimization of the network’s
hyperparameter was applied.
• When contrasted to commonly used one-stage process
frameworks, the two-step framework results in better
accuracy.
5.3.2 Disadvantages
• In instances, the two-step structure compromises speed
or runtime efficiency.
• Does not work well as opposed to other traditional pose
estimation approaches.
5.4 Mask R‑CNN
MaskRCNN is a DNN designed to solve the problem of
instance segmentation in ML or computer vision. Spe-
cifically, it can distinguish between various objects in an
image or video. The fundamental design starts by apply-
ing a CNN to extract feature maps from an image [60]. A
region proposal network (RPN) [24] uses the feature maps to
find bounding box alternatives for the existence of objects.
Because the bounding box candidates can be of different
sizes, a layer named Region of Interest (RoI) Align would
be used to minimize the dimension of the features extracted
such that they were all of the same sizes. The obtained fea-
tures are now forwarded to the parallel branches of CNNs for
the last prediction of the bounding boxes and segmentation
masks. The schematic diagram of Mask RCNN is shown in
Fig.13.
Fig. 12 RMPE Framework [73]

S.Dubey, M.Dixit
1 3
5.4.1 Advantages
• Experience and expertise are not required.
• For each case, creates a segmentation mask.
• Pipeline concurrently anticipates boxes and key points
5.4.2 Disadvantages
• Semantic segmentation can result in erroneous person
bounding boxes, and thus erroneous joint postures.
• Quick, but not velocity optimized
• Leaves no room for a few failure circumstances and
uncommon positions.
5.5 Iterative error feedback
IEF [77] is based on the idea of identifying what is prob-
lematic with the idea of prediction and resolving ititera-
tively, which is accomplished through a top-down feedback
strategy. IEF used a system that integrates both input and
output regions and expands the hierarchical feature extractor
(ConvNet). Input consists of an image I and the originally
predicted keypoints y0 (representing the preceding output
yt-1) can be found on the left side of Fig.13. Consider three
key points: the head (red), right wrist (green), and left wrist
(blue).
Next set input, Xt = I ⊕ g (yt−1),
where, I denote the image and yt-1 is the output from the
previous stage.
The output known as correction εt is produced by the
function f (Xt), which is treated as a ConvNet, and this out-
put is combined with the current output yt to generate yt+1,
which signifies the correction is taken into account t [26].
Each keypointis converted into one Gaussian heatmap
channel using the function g (yt+1), which may then be used
as part of the picture input for the subsequent iteration.
This technique is performed T times until a refined yt+1 is
obtained that is extremely near to the underlying data. This
method assessed their effectiveness on two datasets (LSP
and MPII) using a PCKh@0.5 metric. IEF brought in new
ideas and high-quality work. Both f and g are learnable func-
tions, and f is also a ConvNet. This means that f can learn
features across the joint input–output space.
By considering the system above it can be observed that
deep learning-based approaches have achieved a break-
through in HPE by improving the performance significantly.
5.5.1 Advantages
• The solution is constructed and improved by a series of
steps iteratively. As a result, one can detect flaws in the
preliminary phase.
• Iterative models allocate minimal time to documentation
and more duration to develop.
5.5.2 Disadvantages
• There are no crossovers between the phases of an itera-
tion.
• Since not all needs are obtained up front for the full life-
cycle, there could be costly system architecture or design
concerns.
These are some of the famous deep learning models
which are used for the pose estimation. As we have seen
from the above details each of the model has its own advan-
tages and disadvantages thus, each of these models is better
for the type of application they are applied on.
Fig. 13 Mask RCNN [60]

A comprehensive survey onhuman pose estimation approaches
1 3
6 Occlusion andcrowd detection human
pose estimation
6.1 Occlusion detection andhandling
Occlusion, in which one or even more joints are obscured
from either the camera due to numerous variables such as
self-occlusion, zooming, blurred objects in an image, or
interference by random items, is a major difficulty for human
posture assessment [78]. In a large percentage of real-world
photographs, occlusion happens spontaneously. Occlusion
has been a significant problem in HPE algorithms, particu-
larly those that use deep learning[79]. This problem leads
to inconsistencies in training and testing, which results in
inaccuracy. The ability to handle occlusion, including self-
occlusion [80] and occlusion between people, is a critical
subject in this area. This all depends on the type of data,
including the quality and quantity of the image.
In general, there are three types of occlusions are there in
an image- self-occlusion, inter-object occlusion, and back-
ground occlusion.
6.1.1 Self‑occlusion
In virtual settings, self-occlusion is a difficult challenge to
solve. It arises when a section of an image overlaps the self,
obscuring a part of the image that is ordinarily obscured
from view by another part of the object [81, 82]. It is not that
problematic in static models, but in malleable objects that
are changed in real-time while processing.
6.1.2 Inter‑object occlusion
When the relevant object is obscured by other objects, this is
known as inter-object occlusion. Alike objects, in HPE mul-
tiple persons in crowded environments are difficult to detect
since they are frequently partially or completely obscured
by each other. Therefore, it is very important to handle this
occlusion otherwise it will become one of the major draw-
backs of this field [83].
6.1.3 Background occlusion
When a background object obscures the monitored objects,
this is known as background occlusion (Fig.14).
The review of each of these ways through which occlu-
sion arises using different traditional and deep learning
methods is discussed below and then the challenges and
drawbacks of the suggested methods are observed.
Ghafoor etal. [78], 2022, Occlusion guidance, which
offers additional details regarding the lack or existence of a
joint, is used to address missing joints. Temporal data was
also used to improve the estimation of missing joints. a huge
series of researches are conducted to quantify the presented
method’s occlusion processing capability on three databases
in multiple configurations, such as arbitrary missing joints
and frames.
Qiankun Liu etal. [84], 2022, The suggested occlusion
estimating module aims to forecast where occlusions occur,
which are then utilized to approximate the locations of
objects that are missing. This module tracks multi-object
using unsupervised learning and it is observed from the
results that it is better than the supervised learning method.
The occlusion prediction module can help with the miss-
ingtracking problem.
Renshu Gu etal. [85], 2021, use a gated convolution tem-
poral network to handle the occlusion in 3D pose estima-
tion images. The authors also designed their dataset known
as MMHuman to handle the occluded images in real-world
scenarios.
Yan Di etal. [80], 2021, In this paper authors have over-
come the challenge of directly regressing the 6 Degree
of freedom object pose an RGB image. This can be done
by developing a two-layer model for 3D objects that sig-
nificantly improves end-to-end pose estimation accuracy.
The developed SO-Pose, 6D pose regression framework
Fig. 14 Different ways of self-occlusion [82]

S.Dubey, M.Dixit
1 3
outperforms other single-layer adversaries on a variety of
tough datasets.
Zhou etal. [86], 2020, implemented a Siamese network
for the understanding of occlusion among the different
images. This method has the concept of erasing and recon-
struction i.e., at first erasing step will erase the occluded part
which may sometimes lead to the deletion of the important
features in images therefore, the reconstruction technique is
applied to recover those useful features. Thus, this method
has considered all aspects and implemented a robust model.
Moreover, it has better performance as it has been trained on
a variety of occluded images from different datasets (MPII,
LSP, COCO).
Rohit Kumar Jena [87], 2019 In this research, we present
a new dataset to better assess algorithms and an innovative
and effective way to tackle the issues of crowd pose estima-
tion. They also defined a new dataset based on real-world
applications. The sole major drawback of this work is that it
still does not fully exploit the Part Affinity fields’ potential.
Cheng etal. [88], 2019, to handle occlusion in 3D images
a deep learning-based framework is designed in which par-
tial 2D keypoints are there when occlusion occurs, which
are fed into 2D and 3D CNN. The keypoints selected here
are incomplete as occlusion keypoints errors are less in it.
This network offers a “Cylinder Man Model” to estimate the
utilization of bodily parts in 3D space since no suchdatabase
exists.
Shihui Zhang [81] etal. 2016, This paper proposes a
novel method for estimating movements in 3D space using
depth information of a visual image. To monitor the effi-
ciency of dynamic self-occlusion obfuscation, an assessment
condition called “effective avoidance rate” is introduced.
The experimental findings reveal that the suggested method
meets the purpose of the camera intelligently avoiding self-
occlusion when an object moves (Fig.15).
The paper in the literature shows that despite wide
research in occlusion handling there is still a need for future
research in this field to enhance the reliability and robustness
of the 3D pose estimation. There is still much work required
in handling the 3D pose estimation dataset with occlusion.
Therefore, future researchers should develop an efficient and
simple DL algorithm that handles occlusion easily.
Handling Occlusion- Image segmentation and tracking
of objects are the two approaches that can be used to deal
with occlusions.
6.2 Crowd detection
The problem of crowded scenes in human pose estimation
occurs only while detecting poses in multiperson pose esti-
mation. It is a challenging task because overlaps and occlu-
sions make it much harder to recognize human enclosing
boxes and derive pose indications from particular key points,
MPPE in overcrowded settings is problematic [89]. The
study below shows how to handle the occlusions in crowded
scenarios using various approaches. Not only in HPE, crowd
monitoring is necessary but it also plays a major in many
other applications such as population monitoring, public
event management, suspicious activity detection, military
management, disaster management, and many more[90].
This work presents a bounding box detection and key-
point grouping-free direct pose-level inference technique
known as PINet [89]. It immediately infers a person’s whole
pose cues from his or her observable body components,
rather than suggesting individual key points. This method
is more efficient than other proposed methods but must be
improved to achieve a higher level of accuracy.
Shuning Chang etal. [91], 2020, In this work, the focus
is on enhancing HPE in cluttered scene films. The work is
started by detecting people and doing SPPEusing a top-
down approach is applied for better results.
The results are then processed after applying an algo-
rithm using optical flow which increases the robustness of
the model.
Cheng Chi etal. [92], 2020, The suggested technique,
PedHunter, adds significant occlusion managing capabilities
Fig.15 Inter occlusion detected using different classifiers [83]

A comprehensive survey onhuman pose estimation approaches
1 3
to current region-based detection networks without requiring
additional inference procedures. We create a mask-guided
component that uses head knowledge to optimize the net-
work's function learning algorithm. The algorithm uses
robustness and tests its results on heterogeneous datasets
thus, achieves high accuracy.
Jiefeng Li etal. [93], 2019, In this research, we present
a new dataset for better-evaluating algorithms as well as an
innovative and fast way of tackling the problem of crowd
pose estimation.
A.S. Elons etal.[94], 2017, The research uses a combina-
tion of LSTM and CNN for managing crowded occlusion
thus, using a hybrid deep learning model. The model uses
real-time video frames to analyze and implement a model
with better accuracy.
Extensive studies in the field of crowd analysis have been
conducted recently. There have been numerous datasets
released. But most of these datasets are not able to resolve
the localization and behavior analysis issues of crowd detec-
tion. As a result, there seems to be a lot of room for an in-
crowd assessment of the availabledatabase.
7 Datasets andevaluation metrics
7.1 Dataset
In this section different datasets for 2D/3D pose estima-
tion modeling are elaborated. The content in Tables3 and
4 shows the full description of the 2D and 3D datasets
respectively.
7.2 Performance metrics
Metrics act as an indicator for the pose estimation perfor-
mance [68]. It is difficult to assess the performance of an
estimation of human poses because many factors must be
taken into account [36].
7.2.1 Percentage ofcorrectly estimated body parts (PCP)
PCP evaluates the stick predictions [28]. For a specific part,
PCP can be evaluated as follows,
The higher the value of PCP, the better it works.
(1)
PCP = No
.
of cor r ect par t s for t he entire dat aset
No
.
of t otal parts f or the entire dataset
7.2.2 Percentage ofcorrect key points (PCK)
PCK can be applied in the computation of both 2D and 3D
(PCK3D), in PCK an appendage is viewed as identified (a
right part) if the separation between the two anticipated joint
areas and the actual appendage joint areas is not exactly 50%
of the appendage length (ordinarily meant as PCP at 0.5).
The variation of PCK is PCKh (PCK head) [112]. The limit
of PCKh at 0.5 is half of the head bone band.
7.2.3 Percentage ofdetected joints (PDJ)
If the difference between the expected and the actual artic-
ulation is within a certain part of the torso diameter, the
revealed joint is considered correct [28]. The cons of PDJ
are that it alleviates the shorter limb as the shorter limb has
smaller torsos. PDJ at 0.2 implies that the interval among
expected and actual joints is less than 0.2 * torso diameter.
It is typically used for 2D Pose Estimation. PDJ works better
for the higher values.
7.2.4 Mean perjoint position error (MPJPE)
Mean per joint position error is the average error for all N
joints per joint position (typically, N = 16). The root joints
(typically the pelvis) of the approximate and groundwater-
3D poses are determined after alignment by the method of
similarity transformation. For better understanding, we can
write it as follows,
Per joint position error = calculated Euclidean distance of
ground truth and expected value for a joint.
Average error per joint position = Average error for all N
joints per joint position (typically N = 16).
In MPJPE, the joints ‘J’ is also normalized concerning the
root joint

J
. It can be formalized as follows:
where, M, N is no. of samples and no. of joints
respectively.
MPJPE is mainly used for 3D pose estimation and it gives
better results with lower values.
(2)
PCK at 0.2 =Distance <0.2 ∗
Torso width among expected and true joint
(3)
MPJPE = 1
M
1
N∑M
m=1∑N
n=1‖(
J(m)
n−J
(m)
root
)
−
(

J(m)
n−

J(m)
root
)‖2

S.Dubey, M.Dixit
1 3
Table 3 Description of 2D Datasets used for HPE
Year Dataset name Dataset distribution Type of DATASET Type of physique No. of joints Description Applications
2008 Buffy [95] 472 frames training 276
frames testing
Video Based Upper body 6 body parts Only one person is annotated
in each image
Unconstrained movie and TV
videos
2014 MPII Human Pose Dataset
[96]
Training- 3844
Testing- 1758
Image Based Full Body 16 Keypoints Several interactive people in
strongly articulated positions
with a varying number of
parts are grouped
Identify different activities
such as “Home Activities”,
Garden” Washing windows,
“Picking fruit” or “Rock
climbing”
2010 Leeds Sports Pose [97] Training- 1000
Testing- 1000
Image Based Full body 14 Keypoints Resize images. Only one
person is annotated in each
image
Used to identify the different
types of sports such as “vol-
leyball”, “baseball”, “gym-
nastics”, “parkour” and many
others
2013 Frames Labeled in
Cinema(FLIC) [98]
Training- 3987
Testing- 1016
Image Based Upper body 10 Keypoints Hollywood movies are used to
collect. The people who are
obstructed or significantly
non-frontal are removed
It is used for Human body joint
detection
2014 Frames Labeled in Cinema
(FLIC) Plus [99]
17k images Image Based Upper body 10 Keypoints FLIC-plus is a cleaned and
simple version with no dif-
ficult poses
Enhanced version
2013 BBC Pose [100] Training- 2000
Validation- 1000
Testing- 1000
Video Based Upper body 4 body parts BBC Pose is a collection of
20 videosobtained from the
BBC with a sign language
translator superimposed
Joint Prediction
2017 MSCOCO [101] Training- 118,287
Validation- 5000
Testing-40,670
Image Based Full body 17 Keypoints The data was gathered from
the internet. It has a wide
range of actions
Object detection, segmentation,
and captioning
2017 COCO 17 [102] Training-64k Validation-
2.7k
Testing-40k
Image Based Full body 17 Keypoints Various annotations from
Google, Flickr are taken
Category detection, instance
spotting and instance segmen-
tation
2016 COCO 16 Training-45k Validation-22k
Testing-80k
Image Based Full body 17 Keypoints
2017 PoseTrack [35] Training- 300
Validation- 50
Testing- 208
Video Based Full body 15 Keypoints This dataset focuses on 3
aspects: single-frame MPPE,
MPPE in videos, and articu-
lated tracking
Articulated multiperson pose
tracking, annotated 15 body
parts for each body pose
2014 Parse Training-100
Testing- 205
Image Based Full body 14 Keypoints It's a small dataset with several
annotations, like facial fea-
tures, gender, gaze direction
Focuses on direction of gaze
and facial features

A comprehensive survey onhuman pose estimation approaches
1 3
Table 4 Description of 3D datasets used for HPE
Year Name of dataset Number of images Type of dataset No. of joints Description Applications
2018 3D Poses in the Wild (3DPW)
[103]
60 video sequences
≈ 51k frames
Video Based 18 Keypoints 3DPW is the one to integrate ten
video clips from animations. For
training and testing, the entire
dataset can be exploited
Used to detect future human trajecto-
ries and skeleton poses
2014 Human 3.6M [104] Training≈1.5M Validation≈0.6M
Testing ≈ 1.5M
Video Based 17 Keypoints 3D human representations are cre-
ated through visual effects and
put into complicated real-world
surroundings, which are observed
with movable cameras and in an
indoor space with occlusion
Human activities like talking on
phone, capturing images, posing etc.
It also includes some more synchro-
nized images
2017 MPI-INF-3DHP [105] ≈1.3M Image Based 15 Keypoints This dataset requires 25GB and
7GB spaces for training and test-
ing respectively. Includes indoor
and outdoor scenes
Deep kinematic pose. Good for image
augmentation and segmentation
2018 JTA (Joint Track Auto) [106] Training—256
Testing—256
≈5 00k frames
Video Based – Occlusion Annotation, approx. 10M
body poses
Pedestrian pose estimation and track-
ing in urban scenarios
2017 Total Capture [107] 1,892,176 Image Based
(render from
videos)
26 Keypoints 5 subjects, 5 actions, IMU and Vicon
data, indoor environment
Yoga, walking, giving directions,
bending over, freestyle and crawling
2010 HumanEva I [108] Training ≈ 6.8k Validation≈ 6.8k
Testing ≈ 24k
Video Based 15 Keypoints 3 color + 4 grayscale video cameras,
indoor environment, used for train-
ing, validation, and testing
Jogging, walking, standing, gestur-
ing etc
2010 HumanEva II [109] Testing-2.5k Video Based 15 Keypoints 3color video cameras, mainly used
for testing purposes
2016 TNT15 [110] ≈ 13K frames Video Based 15 Keypoints 4 subjects, 5 actions, IMU data,
indoor environment, 3D body scans
2015 Panoptic [111] 1.5M Video Based 15 Keypoints Multi-annotation), internal environ-
ment
Activity recognition

S.Dubey, M.Dixit
1 3
7.2.5 Object keypoint similarity (OKS)
The difference between expected points and ground truth
points is measured as averaged by an individual’s scale. The
OKS metric is more difficult to compute than the PDJ one.
The OKS can be calculated as,
where, di: toground truth; s: scale the area of the bounding
box divided by the total image area; k: per-key point constant
(4)
OKS = exp(
−d2
i
2s2k2
i
)
that controls fall off; OKS = 1 means that the prediction
done is perfect or you can say that perfect prediction.
7.2.6 Average precision (AP)
Average Precision (AP) metrics are used for the evaluation
of per-frame multi-person pose estimation. This measures
the average recall value of 0 to 1. The other variation of AP
is the Mean Average Precision (MAP) [112].
Here, the below section gives the performance analysis
of different evaluation metrics on the various datasets. The
Tables5, 6, 7, 8, 9 shows the different results based on the
method used.
8 Security attacks
DL and ML models are subject to adversarial instances,
malevolent inputs that have been altered to produce incor-
rect predictive modeling while being unaltered to human
vision [129].
There are various categories of security attacks. One of
them is an attack based on specificity. There are two types
of attacks based on this namely, targeted and untargeted
attacks.
Table 5 Evaluation of 2D HPE methods on LSP and MPII dataset
Higher PCP value on LSP dataset using RNN is 84.2%, Higher PCK
value on LSP dataset using GAN method is 92.1% and on MPII data-
set is 94.0%, and AUC value on LSP dataset using CNN is 65.4 and
on MPII dataset is 61.4 (in bold)
Author Method LSP MPII
PCP
Carreira etal. [77] IEF 72.5% –
Chen etal. [113] Single CNN 75.0% –
Chu etal. [114] Single CNN 81.1% –
Lifshitz etal. [115] RNN 84.2% –
PCK
Bulat etal. [116] Multi-stage CNN 82.7% 83.5%
DeepCut. [75] Single CNN 82.4% 87.1%
Lifshitz etal. [115] RNN 85.0% 85.0%
Belagiannis etal. [68] RNN 88.1% 85.2%
Chu etal. [114] Multi-stage CNN 91.5% 92.6%
Chen etal. [113]GAN 92.1% 93.1%
Chou etal. [117]GAN 91.8% 94.0%
AUC
DeepCut. [75] Single CNN 63.5 56.5
Tompson etal. [99] CNN 47.3 51.8
Carreira etal. [77] IEF 49.1
Wei etal. [29] CNN 65.4 61.4
Table 6 Performance analysis
on Human 3.6M and PoseTrack
dataset
Higher MPJPE value (41.6) on Human 3.6M dataset and Higher average precision of 81.7% on PoseTrack
dataset (in bold)
MPJPE (Average value of all the parameters- eat, greet, phon, pos, pur, sit, wait, walk etc.)
Human 3.6M PoseTrack
Chen etal. [113]GAN 41.6 –
Cai etal. [118] CNN 39.0 –
Wang etal. [119] 32.7 –
Average precision (AP)
Guanghan Ning etal. [120]Ground truth detections – 81.7%
Deform FPN (ResNet101) – 74.6%
Deform R-FCN (ResNet101) – 73.7%
Table 7 Performance analysis on Human 3.0M dataset
Higher S1, S2, S3 and Average value on Human 3.0 M dataset for
Simo Serra etal. [122] (in bold)
References Walking
S1 S2 S3 Avg
Yasin etal. [121] 35.8 32.4 41.6 36.6
Kostrikov and Gall [62] 44.0 30.9 41.7 38.9
Wang etal. [119] 71.9 75.7 85.3 77.6
Simo Serra etal. [122]99.6 108.3 127.4 111.8
Bo and Sminchisescu [123] 38.2 32.8 40.2 37.1

A comprehensive survey onhuman pose estimation approaches
1 3
In a targeted attack [130], there is one targeted class that
has to be misclassified by the attackers on the other hand
there is no specific targeted class to misclassify their main
concern is to misclassify the model with the help of adver-
sarial example [131].
On the basis of the study, it has been analyzed that tar-
geted attacks have high time complexity but are more accu-
rate as compared to untargeted attacks [130, 131].
The other category of adversarial security attack is a
black box and white box attack. In a black Box attack, an
attacker has no prior knowledge of the model in which it is
attacking [132] but in its white box attacks, the attacker has
prior knowledge of the model, and model attributes are eas-
ily accessible to them [133].
Apart from these are other common attacks are,
Integrity attack—A efficient resource assault that is iden-
tified as normal traffic due to false negatives [134].
Availability attack—A broad class of an attack that makes
the system unusable with classification errors, denial of ser-
vice, false negatives and positives, etc. [135].
Privacy violation attack—The discrepancy in confiden-
tial information if not handled properly results in a privacy
violation attack [136].
Hou etal. [137], 2022, the authors have developed a
deep learning-based anomaly detection model known as
the integrity protection method (IPDLS). The main objec-
tive of the method is to determine the feature similarity
between the skeptical sample and the original sample. The
algorithm is evaluated on MNIST and CIFAR10 datasets
and achieved better performance.
Jiawang Bai etal. [130], 2021, To achieve stealth capa-
bilities, the goal is to misidentify a specific instance into
a target class without changing it, but without lowering
the predictive performance of other samples much. The
problem is framed as binary integer programming since
the parameters are stored as bits (i.e., 0 and 1). The defined
technique is resistant to different parameters and is supe-
rior when it comes to attacking DNNs.
Xinghao Yang etal. [135], 2020, here the targeted
attention attack is designed on one of the real-world com-
puter vision applications (road sign recognition attack).
The effective universal assault that optimizes a single per-
turbation based on a collection of training is created with
the help of pre-trained images.
Guowen Xu etal. [138], 2019, the possible risks posed
by deep learning are investigated in this study and the most
up-to-date solutions are determined. Moreover, the authors
have developed the SecureNet which can withstand a vari-
ety of security and privacy risks during a prognosis phase.
Arjun Nitin Bhagoji etal.[134], 2018, The author has
designed the gradient-based black-box attacks without
transferability. In addition to this, methods for decoupling
the number of questions needed to produce each adver-
sarial instance from the input’s dimensionality are also
introduced. The approach is tested on MNIST and CIFAR
10 datasets and achieves a better accuracy,
Yi Shi and Yalin E. Sagduyu [136], 2017, The method
shows that the evasive and causal attacks are initiated
after the exploratory attack, considerably raising the
inaccuracy and also introducing defense techniques. A
defense approach is described that involves changing a
limited number of labels in theactualclassifier to avoid
the adversary from reliably inferring its identity and using
it in evasion and causal assaults. The method is reliable
and takes the varieties of a dataset into consideration (text
and image), thus, able to find out the vulnerabilities in
both types of datasets.
Moustapha Cisse etal. [129], 2017, a versatile method
Houdini is introduced for creating adversarial instances
that are specially customized for the task's final evaluation,
whether combinatorial or non-decomposable. This approach
achieves a significantly higher rate of success in many dif-
ferent areas such as pose estimation, semantic segmentation,
and speech recognition. As a result, the usage of adversarial
instances is extended beyond picture categorization.
Table 8 Evaluation of algorithms for identifying 3D human posture
using various inputs on the MPI-INF-3DHP dataset
Higher PCK, MPJPE and AUC value using DenseRac [126], New DL
based method [125] and RepNet [127] method respectively (in bold)
References Method PCK MPJPE AUC
Zhou etal. [124] Supervised 69.2 – 32.5
Mehta etal. [105] Improved CNN 76.5 – 40.8
Habibie etal. [125] New DL based method 69.6 127.0 35.5
Xu etal. [126] DenseRaC 89.0 83.5 49.1
Wandit and Rosen-
hahn etal. [127]
RepNet 82.5 97.8 58.5
Table 9 Outcomes from the dataset for the MSCOCO keypoint chal-
lenge (AP)
Higher AP value using Faster RCNN with softnms, Pyranet [45] (in
bold)
References Method AP
Cao etal. [128] CMU-Pose 61.8
Papandreou etal. [37] G-RMI 68.5
He etal. [60] Mask R-CNN 63.1
Chen etal. [38] Megvii 72.1
Fang etal. [45] Faster RCNN with softnms,
Pyranet
72.3

S.Dubey, M.Dixit
1 3
9 Conclusion
In this survey, it is observed that as compared to other com-
puter vision problems, the location of parts of the human
body from images and their assembly based on a predefined
human body configuration are involved in estimating the
human pose. By reviewing so many articles it has been
observed that humans and commodities are often considered
isolatedin classic human-object interaction evaluation tech-
niques. The new HPE methodology relies on the interplay
between people and things.
Depending on this, a comprehensive review of various
classical and deep learning 2D/3D single pose and multi-
pose techniques for HPE is studied and it is observed that
each algorithm changes according to its environment and
the availability of the dataset. These techniques and strate-
gies have their benefits and shorts. Furthermore, the vari-
ous human pose datasets for a single person and multi-per-
son 2D/3D have been summed up along with some of the
evaluation metrics. At last, the use of these pose estimation
techniques, and responses to circumstances of body real-
ity are discussed. Finally, it can be concluded that despite
the extraordinary advancement of HPE with deep learning,
there still stay some uncertain difficulties and holes among
research and viable applications, for example, crowd and
occlusion. The most essential concerns for deep learning-
based methodologies are efficient organizations and suffi-
cient training of data. The different security attacks in deep
learning models are also discussed so that the research-
ers should be aware of these attacks and their abnormal
behavior.
9.1 Current andfuture developments
From the study, it can be depicted that the human posse esti-
mation trend is lying towards the deep learning approaches
as it can be noticed that these deep learning techniques
achieve better performance over the activities and the dataset
as compared to other state-of-art approaches. The success
of deep learning approaches to the HPE technique is the
availability of a huge amount of the dataset which is one of
the limitations of the DL application. Despite various data-
bases having been created for unbiased HPE assessments,
additional datasets with adequate examination methodolo-
gies are still desired. Additional body sensors can be used
in the long term to record raw data from diverse postures.
The domain can be separated into two categories: 2D and
3D pose estimate. While 2D PEhas attained an adequate
degree of precision, 3D PE takes a lot of effort unless more
balancing modelsare developed, particularly for interpreta-
tion from a single image and without depth details.
Moreover, the future in HPE is massive and has a great
application area that is important in our daily life. There is
also scope for achieving good results on the higher dimen-
sional datasets (higher than 2D/3D) such as on 6D pose
estimation which estimates the position and direction of the
6D poses. These poses are useful in robotic applications.
Although massive work endeavor efforts are devoted to iden-
tifying human poses from videos or photos, there is indeed
a significant gap between theoretical study and real-world
applications.
Funding Not applicable.
Declarations
Conflict of interest No conflict of interest, financial or otherwise.
Consent for publication Not applicable.
References
1. Chen, Y., Tian, Y., He, M.: Monocular human pose estimation:
a survey of deep learning-based methods. Comput. Vis. Image
Underst. (2020). https:// doi. org/ 10. 1016/j. cviu. 2019. 102897
2. Szczuko, P.: Deep neural networks for human pose estimation
from a very low resolution depth image. Multimed. Tools Appl.
78, 1–21 (2019). https:// doi. org/ 10. 1007/ s11042- 019- 7433-7
3. Liu, Y., Xu, Y., Li, S.: 2-D Human Pose Estimation from Images
Based on Deep Learning: A Review," 2018 2nd IEEE Advanced
Information Management, Communicates, Electronic and Auto-
mation Control Conference (IMCEC), Xi'an, China. 462-465.
(2018). https:// doi. org/ 10. 1109/ IMCEC. 2018. 84695 73
4. Chen, C., Wang, T., Li, D., Hong, J.: Repetitive assembly action
recognition based on object detection and pose estimation. J.
Manuf. Syst. 55, 325–333 (2020). https:// doi. org/ 10. 1016/j. jmsy.
2020. 04. 018
5. Silva, D., Varges, M., Marana, A.N.: "Human action recognition
in videos based on spatiotemporal features and bag-of-poses.
Appl. Soft Comput. 95, 106513 (2020). https:// doi. org/ 10. 1016/j.
asoc. 2020. 106513
6. Ordóñez, F., Roggen, D.: Deep convolutional and LSTM recur-
rent neural networks for multimodal wearable activity recogni-
tion. Sensors 16, 1–25 (2016). https:// doi. org/ 10. 3390/ s1601 0115
7. Christian, S., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.:
Rethinking the inception architecture for computer vision." In
Proceedings of the IEEE conference on computer vision and
pattern recognition. 2818–2826, (2016). https:// doi. org/ 10. 1109/
CVPR. 2016. 308
8. Chen, K., Paolo Gabriel, Alasfour, A., Gong, C., Doyle, W.K.,
Devinsky, O., Friedman, D., etal.: Patient-specific pose estima-
tion in clinical environments. IEEE J. Transl. Eng. Health Med.
6, 1–11 (2018). https:// doi. org/ 10. 1109/ JTEHM. 2018. 28754 64
9. Islam, M.J., Mo J., Sattar. J.: Robot-to-robot relative pose estima-
tion using humans as markers. arXiv preprint arXiv: 1903. 00820
(2019).

A comprehensive survey onhuman pose estimation approaches
1 3
10. Zimmermann, C., Tim, W., Christian, D., Wolfram, B., and
Thomas, B.: 3d human pose estimation in Rgbd images for
robotic task learning. In 2018 IEEE International Conference on
Robotics and Automation (ICRA), pp. 1986–1992. IEEE (2018).
https:// doi. org/ 10. 1109/ ICRA. 2018. 84628 33
11. Vasileiadis, M., Sotiris, M., Dimitrios, G., Christos-Savvas, B.,
Dimitrios, T.: "Robust human pose tracking for realistic service
robot applications." In Proceedings of the IEEE International
Conference on Computer Vision Workshops, pp. 1363–1372.
(2017). https:// doi. org/ 10. 1109/ ICCVW. 2017. 162
12. Preim, B., Meuschke, M.: A survey of medical animations. Com-
put. Graph. 90, 145–168 (2020). https:// doi. org/ 10. 1016/j. cag.
2020. 06. 003
13. Kumarapu, L., Mukherjee, P.: “AnimePose: Multi-person 3D
pose estimation and animation”, arXiv: 2002. 02792 v1, pp 1–5,
(2020). https:// doi. org/ 10. 1016/j. patrec. 2021. 03. 028
14. Tiwari, M.M., Tiwari, M.T., Rajendran, G., Suson, R.: Deep
learning approach for generating 2D pose estimation from video
for motion capture animation. Int. J. Future Gener. Commun.
Netw. 13(2), 1556–1561 (2020)
15. Casado García, F., Luis, Y.,Pérez Losada, D., Santana Alonso,
A.: “Pose estimation and object tracking using 2D images”, In
2017-27th International Conference on Flexible Automation and
Intelligent Manufacturing, Modena, Italy, (2017). https:// doi. org/
10. 1016/j. promfg. 2017. 07. 134
16. Cleetus, A.: Real-time multiple human pose estimation for ani-
mations in game engines. Int. Res. J. Eng. Technol. (IRJET) 7(5),
7923–7928 (2020)
17. https:// mobid ev. biz/ blog/ human- pose- estim ation- ai- perso nal-
fitne ss- coach. Accessed 2 Sept 2021
18. https:// viso. ai/ deep- learn ing/ pose- estim ation- ultim ate- overv iew/.
Accessed 14 Aug 2021
19. Felzenszwalb, P.F., Huttenlocher, D.P.: Pictorial structures for
object recognition. Int. J. Comput. Vision 61, 55–79 (2005).
https:// doi. org/ 10. 1023/B: VISI. 00000 42934. 15159. 49
20. Fischler, M.A., Elschlager, R.A.: The representation and match-
ing of pictorial structures. IEEE Trans. Comput. C22, 67–92
(1973). https:// doi. org/ 10. 1109/T- C. 1973. 223602
21. Yang, Y., Ramanan, D.: Articulated human detection with flex-
ible mixtures of parts. IEEE Trans. Pattern Anal. Mach. Intell.
35(12), 2878–2890 (2013). https:// doi. org/ 10. 1109/ TP AMI. 2012.
261
22. Wu, J., Christopher, G., James M.R.: "Real-time human detection
using contour cues." In2011 IEEE international conference on
robotics and automation, pp. 860–867. IEEE, (2011). https:// doi.
org/ 10. 1109/ ICRA. 2011. 59804 37
23. Micilotta, AS., Eng-Jon, O., Richard, B.: "Real-time upper
body detection and 3D pose estimation in monoscopic images."
In European Conference on Computer Vision, pp. 139–150.
Springer, Berlin, Heidelberg, (2006). https:// doi. org/ 10. 1007/
11744 078_ 11
24. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards
real-time object detection with region proposal networks. IEEE
Trans. Pattern Anal. Mach. Intell. 39(6), 1137–1149 (2016).
https:// doi. org/ 10. 1109/ TPAMI. 2016. 25770 31
25. https:// neura let. com/ artic le/ human- pose- estim ation- with- deep-
learn ing- part-i/. Accessed 15 Sept 2021
26. Munea, T.L., Jembre, Y.Z., Weldegebriel, H.T., Chen, L.,
Huang, C., Yang, C.: The progress of human pose estimation:
a survey and taxonomy of models applied in 2D human pose
estimation. IEEE Access 8, 133330–133348 (2020). https://
doi. org/ 10. 1109/ ACCESS. 2020. 30102 48
27. Rogez, G., Weinzaepfel, P., Schmid, C.: Lcr-net: localization-
classification-regression for human pose. In: Proceedings of
the IEEE Conference on Computer Vision and Pattern Recog-
nition, pp. 3433–3441. (2017)
28. Toshev, A., Szegedy, DC.: Human pose estimation via deep
neural networks, in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, Columbus, WI,
USA, pp. 1653–1660. (2014).
29. Wei, S., Ramakrishna, V., Kanade, T., Sheikh, Y.: "Convolu-
tional Pose Machines," 2016 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA,
pp. 4724–4732. (2016). https:// doi. org/ 10. 1109/ CVPR. 2016.
511
30. Onishi, K., Takiguchi, T., Ariki, Y.: "3D human posture estima-
tion using the HOG features from monocular image." In2008
19th International Conference on Pattern Recognition, pp. 1–4.
IEEE. (2008). DOI:https:// doi. org/ 10. 1109/ ICPR. 2008. 47616
08
31. Newell, A., Yang, K., Deng, J.: Stacked hourglass networks for
human pose estimation, European Conference on Computer
Vision (ECCV) (2016) 483–499 (2016).
32. Bulat, A., Georgios, T.: "Human pose estimation via convolu-
tional part heatmap regression." InEuropean Conference on
Computer Vision, pp. 717–732. Springer, Cham, (2016)
33. Luo, Z., Zhicheng, W., Yan, H., Liang, W., Tieniu, T., Erjin,
Z.. "Rethinking the Heatmap Regression for Bottom-up Human
Pose Estimation." In Proceedings of the IEEE/CVF Confer-
ence on Computer Vision and Pattern Recognition, pp. 13264–
13273. (2021).
34. Chen, X., Yuille, A. L.: Articulated pose estimation by a
graphical model with image dependent pairwise relations,
in Advances in Neural Information Processing Systems, pp.
1736–1744. (2014).
35. Andriluka, M., Iqbal, U., Insafutdinov, E., Pishchulin, L.,
Milan, A., Gall, J., Schiele, B.: “PoseTrack: A Benchmark for
Human Pose Estimation and Tracking”, IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), pp. 5167–
5176. (2018). https:// doi. org/ 10. 1109/ CVPR. 2018. 00542
36. Dang, Qi., Yin, J., Wang, B., Zheng, W.: Deep learning based
2D human pose estimation: a survey. Tsinghua Sci. Technol.
24, 663–676 (2019). https:// doi. org/ 10. 26599/ TST . 2018. 90101
00
37. Papandreou, G., Tyler, Z., Nori, K., Alexander, T., Jonathan,
T., Chris, B., Kevin M.: Towards accurate multi-person pose
estimation in the wild." InProceedings of the IEEE conference
on computer vision and pattern recognition, pp. 4903–4911.
(2017). https:// doi. org/ 10. 1109/ CVPR. 2017. 395
38. Chen, Y., Wang, Z., Peng, Y., Zhang, Z., Yu, G., Sun,
J.:"Cascaded pyramid network for multi-person pose estimation."
InProceedings of the IEEE conference on computer vision and
pattern recognition, pp. 7103–7112. (2018). https:// doi. org/ 10.
1109/ CVPR. 2018. 00742
39. Gamra, M.B., Akhloufi, M.A.: A review of deep learning tech-
niques for 2D and 3D human pose estimation. Image Vis. Com-
put. (2021). https:// doi. org/ 10. 1016/j. imavis. 104282
40. Rodrigues, N., Torres, H.D.R., Oliveira, B., Borges, J., Queirós,
S.F.M., Mendes, J.A., Fonseca, J.C., Coelho, V., Brito, J.H.:
Top-down human pose estimation with depth images and domain
adaptation. SCITEPRESS (2019)
41. Kocabas, M., Karagoz, S., Akbas, E.: "Multiposenet: Fast multi-
person pose estimation using pose residual network." InProceed-
ings of the European conference on computer vision (ECCV), pp.
417–433. (2018). https:// doi. org/ 10. 1007/ 978-3- 030- 01252-6_ 26
42. Insafutdinov, E., Pishchulin, L., Andres, B., Andriluka, M.,
Schiele, B.: "Deepercut: A deeper, stronger, and faster multi-
person pose estimation model." InEuropean Conference on
Computer Vision, pp. 34–50. Springer, Cham, (2016). https://
doi. org/ 10. 1007/ 978-3- 319- 46466-4_3

S.Dubey, M.Dixit
1 3
43. Zheng, C., Wu, W., Yang, T., Zhu, S., Chen, C., Liu, R., Shen, J.,
Kehtarnavaz, N., Shah, M.: "Deep learning-based human pose
estimation: A survey." arXiv preprint arXiv: 2012. 13392 (2020).
44. Cao, Z., Simon, T., Wei, S. E., Sheikh, Y.: "OpenPose:Realtime
multi-person 2d pose estimation using part affinity fields."
InProceedings of the IEEE conference on computer vision and
pattern recognition, pp. 7291–7299. (2017). https:// doi. org/ 10.
1109/ CVPR. 2017. 143
45. Fang, H. S., Xie, S., Tai, Y. W., Lu, C.: "Rmpe: Regional multi-
person pose estimation." InProceedings of the IEEE interna-
tional conference on computer vision, pp. 2334–2343. (2017).
https:// doi. org/ 10. 1109/ ICCV. 2017. 256
46. Tompson, J., Goroshin, R., Jain, A., LeCun, Y., Bregler, C.: "Effi-
cient object localization using convolutional networks." InPro-
ceedings of the IEEE conference on computer vision and pattern
recognition, pp. 648–656. (2015). https:// doi. org/ 10. 1109/ CVPR.
2015. 72986 64
47. Sun, K., Xiao, B., Liu, D., Wang, J.,: "Deep high-resolution rep-
resentation learning for human pose estimation." InProceedings
of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pp. 5693–5703. (2019). https:// doi. org/ 10. 1109/
CVPR. 2019. 00584
48. Osokin, D.: "Real-time 2d multi-person pose estimation on
CPU: Lightweight OpenPose." arXiv preprint arXiv: 1811. 12004
(2018).
49. Tang, W., Yu, P., Wu, Y.: "Deeply learned compositional mod-
els for human pose estimation." InProceedings of the European
conference on computer vision (ECCV), pp. 190–206. (2018).
https:// doi. org/ 10. 1007/ 978-3- 030- 01219-9_ 12
50. Mao, W., Ge, Y., Shen, C., Tian, Z., Wang, X.: "Tfpose: Direct
human pose estimation with transformers." arXiv preprint arXiv:
2103. 15320 (2021).
51. Jain, A., Tompson, J., LeCun, Y., Bregler, C.: "Modeep: A deep
learning framework using motion features for human pose esti-
mation." In: Asian conference on computer vision, pp. 302–315.
Springer, Cham. (2014). https:// doi. org/ 10. 1007/ 978-3- 319-
16808-1_ 21
52. Alzughaibi, A., Chaczko, Z.: "Human detection model using fea-
ture extraction method in video frames,"2016 International Con-
ference on Image and Vision Computing New Zealand (IVCNZ),
pp. 1-6. (2016) https:// doi. org/ 10. 1109/ IVCNZ. 2016. 78044 24
53. https:// mobid ev. biz/ wp- conte nt/ uploa ds/ 2020/ 07/ 3d- keypo ints-
human- pose- estim ation. png. Accessed 20 Aug 2021
54. Hanguen, K., Lee, S., Lee, D., Choi, S., Ju, J., Myung, H.: Real-
time human pose estimation and gesture recognition from depth
images using superpixels and SVM classifier. Sensors (Basel)
(2015). https:// doi. org/ 10. 3390/ s1506 12410
55. Chen, K., Gong, S., Xiang, T.: “Human pose estimation using
structural support vector machines”, 2011 IEEE International
Conference on Computer Vision Workshops (ICCV Workshops),
Barcelona, Spain, pp. 846–851. (2011). https:// doi. org/ 10. 1109/
ICCVW. 2011. 61303 40
56. Hallquist, A., Zakhor, A.: "Single view pose estimation of mobile
devices in urban environments." In2013 IEEE Workshop on
Applications of Computer Vision (WACV), pp. 347–354. IEEE,
(2013).
57. Fei, X., Wang, H., Cheong, L. L., Zeng, X., Wang, M., Tighe, J.:
"Single View Physical Distance Estimation using Human Pose."
InProceedings of the IEEE/CVF International Conference on
Computer Vision, pp. 12406–12416. (2021)
58. Wang, J., Tan, S., Zhen, X., Xu, S., Zheng, F., He, Z., Shao, L.:
Deep 3D human pose estimation: a review. Comput. Vis. Image
Underst. (2021). https:// doi. org/ 10. 1016/j. cviu. 2021. 103225
59. https:// www. kdnug gets. com/. Accessed 30 Aug 2021
60. He, K., Gkioxari, G., Dollár, P., Girshick, R.: "Mask r-cnn."
In Proceedings of the IEEE international conference on
computer vision, pp. 2961–2969. (2017). https:// doi. org/ 10.
1109/ ICCV. 2017. 322
61. Su, J.-Y., Cheng, S.-C., Chang, C.-C., Chen, J.-M.: Model-
based 3D pose estimation of a single rgb image using a deep
viewpoint classification neural network. Appl. Sci. 9(12), 2478
(2019). https:// doi. org/ 10. 3390/ app91 22478
62. Kostrikov, I., Gall, J.: Depth sweep regression forests for esti-
mating 3D human pose from images. BMVC 1(2), 5 (2014).
https:// doi. org/ 10. 5244/C. 28. 80
63. Benzine, A., Chabot, F., Luvison, B., Pham, Q. C., Achard, C.:
"Pandanet: Anchor-based single-shot multi-person 3d pose esti-
mation." InProceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition, pp. 6856–6865. (2020).
64. Mehta, D., Sotnychenko, O., Mueller, F., Xu, W., Sridhar, S.,
Pons-Moll, G., Theobalt, C.: "Single-shot multi-person 3d body
pose estimation from monocular rgb input."arXiv preprint arXiv:
1712. 03453 (2017).
65. Li, S., Chan, A.B.: "3d human pose estimation from monocu-
lar images with deep convolutional neural network." InAsian
Conference on Computer Vision, pp. 332–347. Springer, Cham.
(2014). https:// doi. org/ 10. 1007/ 978-3- 319- 16808-1_ 23
66. Deng, Y., Sun, Y., Zhu, J.: "SVMA: A GAN-based model for
Monocular 3D Human Pose Estimation."arXiv preprint arXiv:
2106. 05616 (2021).
67. Miura, T., Sako, S.: 3D human pose estimation model using loca-
tion-maps for distorted and disconnected images by a wearable
omnidirectional camera. IPSJ Trans. Comput. Vis. Appl. 12(1),
1–17 (2020). https:// doi. org/ 10. 1186/ s41074- 020- 00066-8
68. Belagiannis, V., Amin, S., Andriluka, M., Schiele, B., Navab,
N., Ilic, S., "3D Pictorial Structures for Multiple Human Pose
Estimation," 2014 IEEE Conference on Computer Vision and
Pattern Recognition, Columbus, OH. pp. 1669–1676. (2014).
https:// doi. org/ 10. 1109/ CVPR. 2014. 216
69. Groos, D., Ramampiaro, H., Ihlen, E.A.F.: EfficientPose: scalable
single-person pose estimation. Appl. Intell. 51(4), 2518–2533
(2021). https:// doi. org/ 10. 1186/ s41074- 020- 00066-8
70. Marin-Jimenez, M.J., Romero-Ramirez, F.J., Munoz-Salinas,
R., Medina-Carnicer, R.: 3D human pose estimation from depth
maps using a deep combination of poses. J. Vis. Commun. Image
Represent. 55, 627–639 (2018). https:// doi. org/ 10. 1016/j. jvcir.
2018. 07. 010
71. Pavllo, D., Feichtenhofer, C., Grangier, D., Auli, M.: "3d human
pose estimation in video with temporal convolutions and semi-
supervised training." InProceedings of the IEEE/CVF Confer-
ence on Computer Vision and Pattern Recognition, pp. 7753–
7762. (2019). https:// doi. org/ 10. 1109/ CVPR. 2019. 00794
72. Meng, Lu., Gao, H.: 3D human pose estimation based on a fully
connected neural network with adversarial learning prior knowl-
edge. Front. Phys. 9, 3 (2021). https:// doi. org/ 10. 3389/ fphy. 2021.
629288
73. https:// inblog. in/ Human- Pose- Estim ation- Using- Alpha- Pose-
XyPPE bNTAO. Accessed 10 Sept 2021
74. https:// analy ticsi ndiam ag. com/ guide- to- openp ose- for- real- time-
human- pose- estim ation/. Accessed 10 Sept 2021
75. Pishchulin, L., Insafutdinov, E., Tang, S., Andres, B., Andriluka,
M., Gehler, P. V., Schiele, B.: "Deepcut: Joint subset partition
and labeling for multi person pose estimation." InProceedings of
the IEEE conference on computer vision and pattern recognition,
pp. 4929–4937. (2016). https:// doi. org/ 10. 1109/ CVPR. 2016. 533
76. https:// debug gerca fe. com/ real- time- pose- estim ation- using- alpha
pose- pytor ch- and- deep- learn ing/. Accessed 5 Sept 2021
77. Carreira, J., Agrawal, P., Fragkiadaki, K., Malik, J.: "Human
Pose Estimation with Iterative Error Feedback," 2016 IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), Las Vegas, NV, pp. 4733–4742. (2016). https:// doi.
org/ 10. 1109/ CVPR. 2016. 512

A comprehensive survey onhuman pose estimation approaches
1 3
78. Ghafoor, M., Mahmood, A.: "Quantification of Occlusion Han-
dling Capability of 3D Human Pose Estimation Framework."
IEEE Transactions on Multimedia. (2022). DOI: https:// doi. org/
10. 48550/ arXiv. 2203. 04113
79. Wu, B., Ramakant N.: "Detection of multiple, partially occluded
humans in a single image by bayesian combination of edgelet
part detectors." InTenth IEEE International Conference on
Computer Vision (ICCV'05) Volume 1, vol. 1, pp. 90–97. IEEE,
(2005). https:// doi. org/ 10. 1109/ ICCV. 2005. 74
80. Di, Y., Manhardt, F., Wang, G., Ji, X., Navab, N., Tombari, F.:
"SO-Pose: Exploiting Self-Occlusion for Direct 6D Pose Estima-
tion." InProceedings of the IEEE/CVF International Conference
on Computer Vision, pp. 12396–12405. (2021). DOI: https:// doi.
org/ 10. 1109/ ICCV4 8922. 2021. 01217
81. Zhang, S., He, H., Zhang, Y., Li, X., Sang, Y.: Dynamic self-
occlusion avoidance approach based on the depth image sequence
of moving visual object. Math. Probl. Eng. (2016). https:// doi.
org/ 10. 1155/ 2016/ 47837 94
82. Jacques, J. C., Dihl, L. L., Jung, C. R., Musse, S. R.: "Self-
occlusion and 3D pose estimation in still images." In2013 IEEE
International Conference on Image Processing, pp. 2539–2543.
IEEE. (2013). DOI: https:// doi. org/ 10. 1109/ ICIP. 2013. 67385 23
83. Veld, R. M., Wijnhoven, R. G. J., Bondarev, Y.: "Detection and
handling of occlusion in an object detection system." InVideo
Surveillance and Transportation Imaging Applications 2015, vol.
9407, pp. 184–195. SPIE. (2015). DOI: https:// doi. org/ 10. 1117/
12. 20771 75
84. Liu, Q., Chen, D., Chu, Q., Yuan, L., Liu, B., Zhang, L., Yu, N.:
Online multi-object tracking with unsupervised re-identification
learning and occlusion estimation. Neurocomputing (2022).
https:// doi. org/ 10. 1016/j. neucom. 2022. 01. 008
85. Gu, R., Wang, G., Hwang, J. N.: "Exploring severe occlusion:
multi-person 3d pose estimation with gated convolution." In 2020
25th International Conference on Pattern Recognition (ICPR),
pp. 8243–8250. IEEE. (2021). DOI: https:// doi. org/ 10. 1109/
ICPR4 8806. 2021. 94121 07
86. Antol, S., Lawrence Zitnick, C., Parikh, D.: "Zero-shot learn-
ing via visual abstraction." InEuropean conference on computer
vision, pp. 401–416. Springer, Cham. 2014. https:// doi. org/ 10.
1007/ 978-3- 319- 10593-2_ 27
87. Jena, R.: "Out of the Box: A combined approach for handling
occlusion in Human Pose Estimation."arXiv preprint arXiv:
1904. 11157 (2019).
88. Cheng, Y., Yang, B., Wang, B., Yan, W., Tan, R. T.: "Occlu-
sion-aware networks for 3d human pose estimation in video."
InProceedings of the IEEE/CVF International Conference on
Computer Vision, pp. 723–732. (2019). DOI: https:// doi. org/ 10.
1109/ ICCV. 2019. 00081
89. Wang, D., Zhang, S., Hua, G.: "Robust Pose Estimation in
Crowded Scenes with Direct Pose-Level Inference."Advances
in Neural Information Processing Systems34 (2021).
90. Khan, K., Albattah, W., Khan, R.U., Qamar, A.M., Nayab, D.:
Advances and trends in real time visual crowd analysis. Sensors
(2020). https:// doi. org/ 10. 3390/ s2018 5073
91. Chang, S., Yuan, L., Nie, X., Huang, Z., Zhou, Y., Chen, Y.,
Yan, S.: "Towards accurate human pose estimation in videos of
crowded scenes." InProceedings of the 28th ACM International
Conference on Multimedia, pp. 4630–4634. (2020). DOI: https://
doi. org/ 10. 1145/ 33941 71. 34162 99
92. Chi, C., Zhang, S., Xing, J., Lei, Z., Li, S. Z., Zou, X.: "Ped-
hunter: Occlusion robust pedestrian detector in crowded scenes."
InProceedings of the AAAI Conference on Artificial Intelli-
gence, vol. 34, no. 07, pp. 10639-10646. (2020). DOI: https://
doi. org/ 10. 1609/ AAAI. V34I07. 6690
93. Li, J., Wang, C., Zhu, H., Mao, Y., Fang, H. S., Lu, C.: "Crowd-
pose: Efficient crowded scenes pose estimation and a new
benchmark." InProceedings of the IEEE/CVF conference on
computer vision and pattern recognition, pp. 10863–10872.
(2019). DOI: https:// doi. org/ 10. 1109/ CVPR. 2019. 01112
94. Elons, A.S., Abol-Ela, M.: "Occlusion resolving inside public
crowded scenes based on social deep learning model," 2017
Eighth International Conference on Intelligent Computing and
Information Systems (ICICIS). pp. 218-223. (2017). DOI: https://
doi. org/ 10. 1109/ INTEL CIS. 2017. 82600 50
95. Ferrari, V., Marin-Jimenez, M., Zisserman, A., "Progressive
search space reduction for human pose estimation." In2008 IEEE
Conference on Computer Vision and Pattern Recognition, pp.
1–8. IEEE. (2008). https:// doi. org/ 10. 1109/ CVPR. 2008. 45874 68
96. Shafaei, A., James J.L.: “Real-Time Human Motion Capture with
Multiple Depth Cameras”, Proceedings of the 13th Conference
on Computer and Robot Vision. (2016). https:// doi. org/ 10. 1109/
CRV. 2016. 25
97. Johnson, S., Everingham, M.: “Learning Effective Human Pose
Estimation from Inaccurate Annotation”, Proceedings of IEEE
Conference on Computer Vision and Pattern Recognition, pp
1465–1472. (2011). https:// doi. org/ 10. 1109/ CVPR. 2011. 59953
18
98. Sapp, B., Taskar, B.: “MODEC: Multimodal Decomposable
Models for Human Pose Estimation”, In 2013 IEEE Conference
on Computer Vision and Pattern Recognition, NW Washington
DC, United States, pp. 3674–3681. (2013). https:// doi. org/ 10.
1109/ CVPR. 2013. 471
99. Tompson, J.J., Jain, A., LeCun, Y., Bregler, C.: Joint training
of a convolutional network and a graphical model for human
pose estimation. Adv. Neural. Inf. Process. Syst. 27, 1799–1807
(2014)
100. Charles, J., Pfister, T., Everingham, M., Zisserman, A.: Auto-
matic and efficient human pose estimation for sign language
videos. Int. J. Comput. Vision 110(1), 70–90 (2014). https:// doi.
org/ 10. 1007/ s11263- 013- 0672-6
101. T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan,
P. Dollar, and C. L. Zitnick, Microsoft coco: Common objects in
context, in European Conference on Computer Vision, Zurich,
Switzerland, 2014, pp. 740–755.
102. Bin, Y., Chen, Z. M., Wei, X. S., Chen, X., Gao, C., Sang, N.:
“Structure-aware Human Pose Estimation with Graph Convo-
lutional Networks”, Vol. 106, pp.107410, Pattern Recognition.
(2020). https:// doi. org/ 10. 1016/j. patcog. 2020. 107410
103. Von Marcard, T., Henschel, R., Black, M. J., Rosenhahn, B.,
Pons-Moll, G.: "Recovering accurate 3d human pose in the wild
using imus and a moving camera." InProceedings of the Euro-
pean Conference on Computer Vision (ECCV), pp. 601–617.
(2018).
104. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.:
Human3.6M: Large Scale Datasets and Predictive Methods for
3D Human Sensing in Natural Environments, IEEE Transactions
on Pattern Analysis and Machine Intelligence, vol. 36, No. 7.
(2014). https:// doi. org/ 10. 1109/ TPAMI. 2013. 248
105. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu,
W., Theobalt, C.: "Monocular 3d human pose estimation in the
wild using improved cnn supervision." In2017 international con-
ference on 3D vision (3DV), pp. 506–516. IEEE. (2017). https://
doi. org/ 10. 1109/ 3DV. 2017. 00064
106. Fabbri, M., Lanzi, F., Calderara, S., Palazzi, A., Vezzani, R.,
Cucchiara, R.: "Learning to detect and track visible and occluded
body joints in a virtual world." In Proceedings of the European
conference on computer vision (ECCV), pp. 430–446. (2018).
https:// doi. org/ 10. 1007/ 978-3- 030- 01225-0_ 27
107. Trumble, M., Gilbert, A., Malleson, C., Hilton, A., Collomosse,
J.P.: Total capture: 3D human pose estimation fusing video and
inertial sensors. BMVC 2(5), 1–13 (2017). https:// doi. org/ 10.
5244/C. 31. 14

S.Dubey, M.Dixit
1 3
108. Sigal, L., Balan, A., Black, M.J.: HumanEva: synchronized video
and motion capture dataset and baseline algorithm for evaluation
of articulated human motion. Int. J. Comput. Vis. 87, 1–2 (2010)
109. Sigal, L., Black, M. J.: HumanEva: Synchronized video and
motion capture dataset for evaluation of articulated human
motion,Techniacl Report CS-06–08, Brown University. (2006).
110. Marcard, T.V., Pons-Moll, G., Rosenhahn, B.: “Multimodal
motion capture dataset TNT15”. Leibniz Univ. Hannover, Hano-
ver, Germany, and Max Planck for Intelligent Systems, Tübingen,
Germany. Tech. Rep. (2016). https:// doi. org/ 10. 13140/ RG.2. 1.
4162. 0248
111. Joo, H., Liu, H., Tan, L., Gui, L., Nabbe, B., Matthews, I.,
Kanade, T., Nobuhara, S., Sheikh, Y.: "Panoptic studio: A mas-
sively multiview system for social motion capture." InProceed-
ings of the IEEE International Conference on Computer Vision,
pp. 3334–3342. 2015. DOI:https:// doi. org/ 10. 1109/ ICCV. 2015.
381
112. Haque, A., Peng, B., Luo, Z., Alahi, A., Yeung, S., Fei-Fei, L.:
“Towards Viewpoint Invariant 3D Human Pose Estimation”,
arXiv: 1603. 07076 v3 [cs.CV], pp.1–20. (2016). https:// doi. org/
10. 1007/ 978-3- 319- 46448-0_ 10
113. Chen, Y., Shen, C., Wei, X. S., Liu, L., Yang, J.: "Adversarial
posenet: A structure-aware convolutional network for human
pose estimation." InProceedings of the IEEE International
Conference on Computer Vision, pp. 1212–1221. (2017). DOI:
https:// doi. org/ 10. 48550/ arXiv. 1705. 00389
114. Chu, X., Yang, W., Ouyang, W., Ma, C., Yuille, A.L., Wang,
X.: Multi-context attention for human pose estimation. CVPR
(2017). https:// doi. org/ 10. 48550/ arXiv. 1702. 07432
115. Lifshitz, I., Fetaya, E., Ullman, S.: Human pose estimation using
deep consensus voting. ECCV (2016). https:// doi. org/ 10. 1007/
978-3- 319- 46475-6_ 16
116. Bulat, A., Tzimiropoulos, G.: Human pose estimation via convo-
lutional part heatmap regression. ECCV (2016). https:// doi. org/
10. 1007/ 978-3- 319- 46478-7_ 44
117. Chou, C. J., Chien, J. T., Chen, H. T.: "Self adversarial training
for human pose estimation." In2018 Asia-Pacific Signal and
Information Processing Association Annual Summit and Confer-
ence (APSIPA ASC), pp. 17–30. IEEE. (2018). DOI: https:// doi.
org/ 10. 48550/ arXiv. 1707. 02439
118. Cai, Y., Ge, L., Liu, J., Cai, J., Cham, T.-J., Yuan, J., etal.:
Exploiting spatial-temporal relationships for 3D pose estimation
via graph convolutional networks. In: International conference
on computer vision p. 227–2281. (2019). doi: https:// doi. org/ 10.
1109/ ICCV. 2019. 00236 48.
119. Wang, J., Yan, S., Xiong, Y., Lin, D.: "Motion guided 3d pose
estimation from videos." InEuropean Conference on Computer
Vision, pp. 764–780. Springer, Cham. (2020). DOI: https:// doi.
org/ 10. 48550/ arXiv. 2004. 13985
120. Ning, G., Liu, P., Fan, X., Zhang, C.: "A top-down approach
to articulated human pose estimation and tracking." InProceed-
ings of the European Conference on Computer Vision (ECCV)
Workshops, pp. 0–0. (2018). DOI: https:// doi. org/ 10. 1007/ 978-
3- 030- 11012-3_ 20
121. Yasin, H., Iqbal, U., Kruger, B., Weber, A., Gall, J.: "A dual-
source approach for 3d pose estimation from a single image."
InProceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 4948–4956. (2016). DOI: https:// doi.
org/ 10. 1016/j. cviu. 2018. 03. 007.
122. Simo-Serra, E., Quattoni, A., Torras, C., Moreno-Noguer, F.: "A
joint model for 2d and 3d pose estimation from a single image."
InProceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 3634–3641. (2013). DOI: https:// doi.
org/ 10. 1109/ CVPR. 2013. 466
123. Bo, L., Sminchisescu, C., Kanaujia, A., Metaxas, D.: "Fast algo-
rithms for large scale conditional 3D prediction." In2008 IEEE
Conference on Computer Vision and Pattern Recognition, pp.
1–8. IEEE. (2008). DOI: https:// doi. org/ 10. 1109/ CVPR. 2008.
45875 78
124. Zhou, X., Huang, Q., Sun, X., Xue, X., Wei, Y.: Towards 3d
human pose estimation in the wild: a weakly-supervised
approach. In: Proceedings of the IEEE International Conference
on Computer Vision. pp. 398–407. (2017). DOI: https:// doi. org/
10. 48550/ arXiv. 1704. 02447
125. Habibie, I., Xu, W., Mehta, D., Pons-Moll, G., Theobalt, C.: In
the wild human pose estimation using explicit 2D features and
intermediate 3D representations. In: Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition. pp.
10905–10914. (2019). DOI: https:// doi. org/ 10. 48550/ arXiv . 1904.
03289
126. Xu, Y., Zhu, S.-C., Tung, T.: DenseRaC: Joint 3D Pose and Shape
Estimation by Dense Render-and-Compare. In: Proceedings of
the IEEE International Conference on Computer Vision. pp.
7760–7770. (2019). DOI: https:// doi. org/ 10. 48550/ arXiv. 1910.
00116
127. Wandt, B., Rosenhahn, B.: RepNet: Weakly supervised training
of an adversarial reprojection network for 3D human pose esti-
mation. In: Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition. pp. 7782–7791. (2019). DOI:
https:// doi. org/ 10. 48550/ arXiv. 1902. 09868
128. Chen, X., Lin, K., Liu, W., Qian, C., Lin, L.: Weakly-supervised
discovery of geometry-aware representation for 3D human pose
estimation. In: Conference on computer vision and pattern rec-
ognition p. 10895–904. (2019).
129. Cisse, M. M., Adi, Y., Neverova, N., Keshet, J.: Houdini: Fooling
deep structured visual and speech recognition models with adver-
sarial examples. In: Advances in neural information processing
systems, vol. 30. (2017)
130. Bai, J., Wu, B., Zhang, Y., Li, Y., Li, Z., Xia, S. T.: "Targeted
attack against deep neural networks via flipping limited weight
bits."arXiv preprint arXiv: 2102. 10496. (2021).
131. Rathore, P., Basak, A., Nistala, S. H., Runkana, V.: "Untargeted,
Targeted and Universal Adversarial Attacks and Defenses on
Time Series." In2020 International Joint Conference on Neural
Networks (IJCNN), pp. 1–8. IEEE. (2020). DOI: https:// doi. org/
10. 1109/ IJCNN 48605. 2020. 92072 72
132. Guo, S., Zhao, J., Li, X., Duan, J., Mu, D., Jing, X.: A black-box
attack method against machine-learning-based anomaly network
flow detection models. Secur. Commun. Netw. (2021). https://
doi. org/ 10. 1155/ 2021/ 55783 35
133. Wang, Y., Liu, J., Chang, X., Wang, J., Rodríguez, R. J.: "DI-
AA: An Interpretable White-box Attack for Fooling Deep Neural
Networks."arXiv preprint arXiv: 2110. 07305. (2021).
134. Bhagoji, A. N., He, W., Li, B., Song, D.: "Exploring the space
of black-box attacks on deep neural networks." arXiv preprint
arXiv: 1712. 09491. (2017).
135. Yang, X., Liu, W., Zhang, S., Liu, W., Tao, D.: Targeted attention
attack on deep learning models in road sign recognition. IEEE
Internet Things J. 8(6), 4980–4990 (2021). https:// doi. org/ 10.
1109/ JIOT. 2020. 30348 99
136. Shi, Y., Sagduyu, Y. E.: "Evasion and causative attacks with
adversarial deep learning,"MILCOM 2017 - 2017 IEEE Military
Communications Conference (MILCOM). pp. 243–248. (2017).
doi: https:// doi. org/ 10. 1109/ MILCOM. 2017. 81708 07.

A comprehensive survey onhuman pose estimation approaches
1 3
137. Hou, R., Ai, S., Chen, Q., Yan, H., Huang, T., Chen, K.: Simi-
larity-based integrity protection for deep learning systems. Inf.
Sci. (2022). https:// doi. org/ 10. 1016/j. ins. 2022. 04. 003
138. Xu, G., Li, H., Ren, H., Yang, K., Deng, R.H.: Data security
issues in deep learning: attacks, countermeasures, and opportuni-
ties. IEEE Commun. Mag. 57(11), 116–122 (2019). https:// doi.
org/ 10. 1109/ MCOM. 001. 19000 91
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.