Available via license: CC BY 4.0
Content may be subject to copyright.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
https://doi.org/10.62617/mcb468
1
Article
Application of human-computer interaction technology integrating
biomimetic vision system in animation design with a biomechanical
perspective
Jing Han
School of Art and Design, Yantai Institute of Science and Technology, Yantai 265600, China; 13081635526@163.com
Abstract: The combination of Human-Computer Interaction (HCI) technology with
biomimetic vision systems has transformational potential in animation design, particularly by
incorporating biomechanical principles to create immersive and interactive experiences.
Traditional animation approaches frequently lack sensitivity to real-time human motions,
which can restrict engagement and realism. This study addresses this constraint by creating a
framework that uses Virtual Reality (VR) and Augmented Reality (AR) to generate dynamic
settings that include a variety of human activities, informed by biomechanical analysis. A
biomimetic vision system is used to record these motions with wearable sensors, allowing for
precise monitoring of user activity while considering biomechanical factors such as joint angles,
force distribution, and movement patterns. The recorded data is preprocessed using Z-score
normalization methods and extracted using Principal Component Analysis (PCA). This study
proposed an Egyptian Vulture optimized Adjustable Long Short-Term Memory Network
(EVO-ALSTM) technique for motion classification, specifically tailored to recognize
biomechanical characteristics of human movements. Results demonstrate a significant
improvement in precision (93%), F1-score (91%), accuracy (95%), and recall (90%) for the
motion recognition system, highlighting the effectiveness of biomechanical insights in
enhancing animation design. The findings indicate that integrating real-time biomechanical
data into the animation process leads to more engaging and realistic user experiences. This
study not only advances the subject of HCI but also provides the framework for future
investigations into sophisticated animation technologies that use biomimetic and
biomechanical systems.
Keywords: animation; motion recognition; biomimetic vision system; human activities;
biomechanics; Human-Computer Interaction (HCI)
1. Introduction
Human-computer interaction (HCI) technology plays an important function in
animation design since it allows natural interaction between people and computers. It
permits animators to create and manipulate complicated visual elements through user-
friendly interfaces and interactive systems [1]. When applied in the layout of
animation, the approach is additionally efficient, inventive, and accessible to designers,
enabling them to convey their ideas into consequence more accurately. It reduces the
need for lots of manual intervention and narrows down the creative process. It presents
new opportunities for launching innovative and complex animation projects [2].
A biomimetic vision system is an artificial visible processing generation that
mimics biological vision, mainly human or animal vision structures. These systems
are designed to capture, process, and interpret visual information further to how
CITATION
Han J. Application of human-
computer interaction technology
integrating biomimetic vision system
in animation design with a
biomechanical perspective.
Molecular & Cellular Biomechanics.
2024; 21(4): 468.
https://doi.org/10.62617/mcb468
ARTICLE INFO
Received: 8 October 2024
Accepted: 23 October 2024
Available online: 25 December 2024
COPYRIGHT
Copyright © 2024 by author(s).
Molecular & Cellular Biomechanics
is published by Sin-Chn Scientific
Press Pte. Ltd. This work is licensed
under the Creative Commons
Attribution (CC BY) license.
https://creativecommons.org/licenses/
by/4.0/
Molecular & Cellular Biomechanics 2024, 21(4), 468.
2
residing organisms work [3]. In animation design, incorporating a biomimetic vision
system allows for more realistic animations because the system can observe and
interpret natural movements and visual cues, ensuring accurate person and
surroundings rendering. It also adapts to varying lighting and environmental situations
to reflect how human vision works in diverse settings [4]. The era can help designers
obtain better ranges of elements and visual accuracy in their work.
The integration of the biomimetic vision framework into animation enhances the
realism of the visual factors. By simulating the way human beings understand light
intensity, color, and movement, this era permits animators to create reasonable scenes
and characters [5]. It results in extra immersive and convincing animators, where the
movers interact with the surrounding experiences naturally, enriching the viewer’s
experience. Realism can help animations evoke stronger emotional responses from the
target audience [6]. Moreover, it lets in for extra attractive storytelling through making
the visuals seem towards real life.
HCI technology, when mixed with biomimetic vision, can massively enhance the
interactive design of animations. Animators and architects can engage with the system
more intuitively through the use of gestures, eye tracking, or voice instructions to
control the animation technique [7]. The hand-free approach allows for smoother
workflows and extra engaging user experiences, making it less complicated to create
dynamic animations that reply to real-time inputs. Additionally, it fosters an extra
immersive improvement where designers can, at once, affect results with minimum. It
ensures greater innovative flexibility and reduces repetitive guide responsibilities.
In 3D animation and visible outcomes, the biomimetic vision system can simulate
complicated visual phenomena inclusive of reflections, shadows, and textures with
higher accuracy [8]. These structures analyze the scene as a human eye track,
supporting animators to generate more specified and visually attractive environments.
The technology can also enhance motion capture using imitation of how human eyes
follow movement, resulting in additional natural and fluid character movements. Thus,
biomimetic vision makes it easy to grab and record subtle information, including
changes in texture or light, to produce near-perfect digital worlds [9]. It contributes a
dimension of professional excellence to tasks, generally making them even more
pleasant across the board.
For animation designers and manufacturers, the combination of HCI technology
and biomimetic vision systems offers significant benefits. The time and effort exerted
in the production of quality animation are also cut down, in that the device optimizes
the critical animated values mainly based on real-world recorders [10]. Furthermore,
this era also allows designers to explicitly express more visible styles and concepts
while enabling realistic and interactive designs. It additionally makes collaboration
simpler, as a couple of designers can work seamlessly with the device. It leads to
quicker manufacturing cycles and complements the overall quality [11].
To enhance animation design, the integration of the HCI era and biomimetic and
prescient systems can be applied to human activity. By leveraging actual time motion
tracking and visual recognition, those technologies can analyze and interpret human
moves with excessive precision. Biomimetic vision systems mimic the way human
movements, bearing in mind more accurate classification of various activities together
[12]. The approach is particularly useful in fields like animation, digital reality, and
Molecular & Cellular Biomechanics 2024, 21(4), 468.
3
video game development, wherein sensible character moves are crucial. Figure 1
represents the general framework of virtual interaction and human activities.
Figure 1. The general framework of virtual interaction and human activities.
HCI generation additionally complements the process by supplying intuitive
interplay methods, including gesture control and body movement monitoring,
permitting designers to seamlessly manipulate animations that accurately replicate
human activities. Additionally, device mastering algorithms can be integrated to
continuously improve the accuracy of interest type by getting to know big information
about human motion [13]. The effects include smoother and more fluid animations
that reflect human movements; this integration streamlines workflows complements
realism, and opens new opportunities for interactive actual-time animation
improvement.
The paper aims to improve animation design by proposing an Egyptian vulture-
optimized adjustable long short-term memory network (EVO-ALSTM) for human
activity classification.
Key contributions
• The data was gathered from 30 participants using wearable sensors, which
included accelerometers and gyroscopes.
• Z-score normalization was used for preprocessing and PCA was used to extract
the complex features from the preprocessed data.
• EVO-ALSTM is proposed to classify human activity.
The remaining parts of this paper: Part 2 represents the related work, a
methodology that includes dataset, preprocessing, and feature extraction, and the
proposed method was described in Part 3. Part 4 presents the result and discussion.
Part 5 covered the paper’s conclusion.
2. Related work
To enhance the natural relationship between humans and machines, [14] created
a detailed blueprint for a robot with a humanoid head that possessed human-like
feelings and activities. It also evaluated how well human behavior and emotional
Molecular & Cellular Biomechanics 2024, 21(4), 468.
4
expression were simulated by the motor and sensory control system. The field of
biomimetic humanoid robots benefited from these efforts. Experimental data from the
survey indicated that participants expressed emotions, and participants replicated
actions.
A didactic, graphical tool that presented the most recent applications of
biomimicry in medicine [15] suggested a webpage was created to show 2D artwork
and visuals (animations). Animation was a tried-and-true method of educating the
public about health-related topics. All respondents agreed that biomimicry could
provide useful responses for medical design. Investigation showed that for public
outreach, visual motions could successfully communicate complex ideas.
A presentation on 2D animation using artificial intelligence and biomechanics
modeling (2D-AI-BM) was presented in [16]. Deep neural networks (DNN) for
movement predictions and development based on biopsychological principles were
employed in the process to better resemble real human motions. Research contrasting
that approach with conventional animation approaches has demonstrated that it
reduced the production period while indeed generating realistic movements for 2D
characters. It presented numerical results proving that the implementation of the 2D-
AI-BM model enhances an accuracy rate.
A flexible simulated tactile approach that made use of the stick-slip sensing
model in [17] offered a general approach to identify a failure and quantify the surface
properties of an object through slippage. The system comprised a read-out system in
the form of tips of hands, a display unit, and an artificial intelligence component.
Based on the stick-slip sensing approach, the system had a high identification rate for
slippage monitoring. The multipurpose system was also demonstrated for interactive
gaming, robotic hand deception, and identifying objects, allowing for extensive and
prospective interactions between humans and machines.
A new architecture that employed motion data recorded [18] by human webcams.
Because of that technique, which used a lot of movement information in the real-time
recreation of such animation as animals moving, designers could be in a position to
design those characters with more accurate movements, which could depict real-world
settings. Moreover, users’ actions were tied to virtual reality, making the whole action
more realistic and exciting.
A model of random forests [19] was to process and create animation data, and
from the collected animation data, the knowledge that could guide the development of
animation was extracted. Based on the design goal and execution approach of the
animated information processing and development platform, the features and
categories of the random forest model were separated. The findings from the
experiment showed that the platform for developing and processing 3D animation data
was both practical and efficient.
Created several techniques to expedite and reduce the expense of that procedure
[20] developing a mobile robot that could follow the actor and record the scene while
keeping them where he needs to be in the frame. By feeding the recorded video
through a range of deep learning algorithms, the team could then determine the actor’s
3D position. It could be used to animate the required 3D model; therefore, there is no
need to use several cameras and a mobcap suit to capture the movements of the actor.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
5
A customized DNN to constantly and accurately detect external haptic stimuli
[21] suggested a new method of data enhancement process was pioneered by
identifying the hexagonal structure of the sensor, which has six-fold rotational
symmetry and possesses mirror images. The generated pseudo data could enhance the
generalization performance of the DNN model by adding the obtained training data.
The sensor proved its effectiveness and the feasibility of the proposed data
augmentation technique and provided a good generalization of five touch modes and
potential for further development to improve human-robot interaction.
Four methods for incorporating deep learning models and Kinect camera-based
animated manufacturing systems with natural human movement [22] were examined.
The selection of each approach was contingent upon environmental circumstances and
accuracy. The initial solution made use of a Kinect camera. A camera and a calibration
algorithm were employed in the second technique. The third option made use of a deep
learning framework. A deep learning model was employed when combined with
Kinect in the fourth method. Comparing the recommended method’s experiments to
previous approaches, it was found that the fourth method, which combines a Kinect
and a deep learning model, produced the greatest results.
A complex HCI program to point at the problem of communication impairment
between groups of individuals with hearing disabilities and without such problems was
designed [23]. The advancement of artificial intelligence has made it quite possible by
the hard work of listening and people without disabilities to communicate as per their
desire. Using near-wearable technology and utilizing backpropagation (BP) neural
network models to classify gestures, the proposed system was able to successfully
close the communication gap between the impaired and the non-impaired individuals.
A revolutionary biomimetic bidirectional cooperation perceiving system
(BBCPS) [24] suggested the gaze function in human eyes served as inspiration for
their creation of a simple yet versatile BBCPS. The results of the simulation
demonstrate that it reduced operational energy consumption and enhanced braking
effectiveness. Furthermore, a system for initial position calibration was established to
ensure that the BBCPD state matches the control strategy that followed. The
methodology allowed for the certification and modification of the camera pose and
servo motors’ zero-position. The gaze error was fewer than three pixels across in real
testing, confirming the control performance of the BBCPS.
Neurological control systems [25] enhance significantly the realism of the
simulation of human movement. They have underscored some of the challenges in
relocating head pose and facial emotions from the pictures and clinical movies into
muscle-actuated modeling of facial and head and neck. A complicated biomechanical
system was also involved in generating locomotion-based animations with
biomechanical plausibility. It extended the use of imitation, physics-based humanoid
simulators, and modeling in graphic design and vision by showcasing the adaptability
of the face and body controlled by muscles.
A unique learning-based method for biomechanically modeling the face-head-
neck complicated by importing facial emotions and head motions from images and
videos was described in [26] and suggests training a DNN to take in the face actions
coding system (FACS) with action units (AUs) and produce appropriate facial muscle
and mouth motion signals for the biomechanics model by using the FACS as a
Molecular & Cellular Biomechanics 2024, 21(4), 468.
6
substitute for representing emotion distance. Experiments including the projection of
different facial emotions and head poses from films onto the face-head-neck model
demonstrated the model’s efficacy.
A simple yet efficient method that used deep learning (DL) [27] to create a basic
3D animation of numerous people moving in 2D. Despite recent considerable
advances in 3D human posture calculation, multi-person determination of pose was
rather a challenging topic, and most previous works were yet limited to single-person
estimation of poses. Using the publicly available dataset, the proposed system
performed comparably to prior state-of-the-art 3D multi-person pose approximation
approaches and surpassed previous competitive human pose tracking devices by a
significant margin.
The efficacy and precision of human annotators, whether employing video, data,
or both for annotating events across four human activity recognition (HAR) tasks [28]
observed that annotators were more accurate in classifying kinds of events when
employing video alone on all four tasks and more effective while using data alone on
three of the four assignments. The annotations of event boundaries based on data alone
were more accurate. The experimental findings discovered that the data and video
collected for HAR task annotations had multiple functions and that the functions might
vary across the HAR tasks.
A novel system for deep learning based on signals from movement to identify
human activities and address these limitations and difficulties using deep learning
techniques [29] Utilized convolutional neural networks (CNNs) and laboratory
metrics, the methodology was effectively studied and obtained better accuracy in
comparison to machine learning techniques. The research’s innovative approach was
to improve classification accuracy while executing tasks more quickly and with a
lower mistake rate. It also introduced a new technique that uses CNN with Adam’s
optimization technique to detect human involvement in the dataset.
Problem statement
The current animation design using HCI with a biomimetic vision system brings
massive challenges, such as capturing and responding to actual time human motion.
Traditional animation techniques regularly fail to offer the level of sensitivity and
interactivity required for attractive and practical user experiences. Existing
frameworks often struggle to appropriately interpret complicated moves, leading to a
disconnection between user actions and animated responses. There is also a lack of
standardized protocols for integrating HCI and biomimetic systems, complicating the
development of cohesive frameworks that make certain compatibility and
effectiveness. The high cost related to the advanced technology implementation can
restrict accessibility for big adoption and innovation within the field. These challenges
collectively hinder the potential of animation design to create trust immersive and
interactive investigations that resonate with users. The proposed method deals with
the limitations of cutting-edge animation techniques with the aid of growing a
framework that utilizes VR and AR to create dynamic environments that contain
numerous human activities.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
7
3. Methodology
The data consists of analysis of the relevant data where Z-score normalization is
needed to standardize the distribution of the data. To reduce dimensionality, and retain
the most important variance, Principal Component Analysis (PCA) is used. The
present method EVO-ALSTM, combines the benefits of Egyptian Vulture optimized
adjustable Long Short-Term Memory networks prediction power as well as optimizing
the learning process adaptively. The purpose of the coupling is to focus on improving
the performance of the model concerning complex data sets these overall procedures
are shown in Figure 2.
Figure 2. Methodology’s overall procedure.
3.1. Dataset
The dataset was gathered from Kaggle
(https://www.kaggle.com/datasets/ziya07/human-motion-dataset-for-animation-
design/data). The dataset comprises human activity data acquired using attached
motion sensors; accelerometers and gyroscopes, which were embedded into a
biomimetic vision sensor system on thirty subjects (15-males, 15-females). This
system intraocular visualizes dynamic activities as a human eye would increase the
precision in such activities. Each of the subjects performed a set of defined motions,
namely walking, jumping, waving arms, and performing sports-mimicking actions.
The activities were recorded in an enclosed area as most of the activities were captured
and implemented in a 3D world so a 3D vision system was used along with augmented
reality (AR). Before data gathering, the sensors and the imaging system were set up to
enhance their functional efficiency and accuracy for the measurements of motion and
body dynamics. Each session per participant lasted 30 min with the directive to act as
freely as possible while the recordings were taken. The data obtained, which was
further enhanced by the imaging system, was collected and classified in an orderly
Molecular & Cellular Biomechanics 2024, 21(4), 468.
8
fashion along the lines of subject identification number and type of activity to facilitate
in-depth tracking of human movements for other purposes.
3.2. Z score normalization
The process of normalizing data involves scaling or mapping abnormal data to
standard data. The Z-score approach, a numerical data type, is used in the model and
its normalized values in the dataset. A common statistical method for standardizing
and normalizing numerical features in a dataset is the Z-score method. It calculates the
Z-score for every fact point by subtracting the implied and dividing, utilizing the
standard deviation of the dataset. The normalization statistics can have a mean of 0
and a popular deviation of 1, improving the model’s overall performance in the
movement class. It facilitates minimizing the influence of outliers and ensures that
different capabilities contribute similarly to the type system. The formula for Z-score
normalization is represented in Equation (1).
µ
(1)
3.3. Extraction of feature
PCA utilized to remove features to maximize records variability and then convert
it directly into a space with a low number of dimensions. It is a powerful method used
for feature extraction in movement category responsibilities related to human activity
datasets. By reducing the dimensionality of the statistics, PCA identifies the most
sizeable capabilities that account for the variance in human actions, facilitating
improved model performance. The method eliminates the noise and redundant data,
making it easier to classify critical patterns and movements. This algorithm can
function extra efficaciously and correctly, enhancing the recognition of numerous
human activities and it contributes to extracting effective and reliable movement
classification structures. An ortho basis set that is identical is the resultant vector set.
Since the fundamental elements are the vectors that form part of the balanced
interaction matrix, each of them is orthogonal. In mathematics, if samples are taken
from a dataset and the class label is not considered, then every measurement is n-
dimensional. Assume ℜ that the following steps for PCA
calculation.
Determine the mean vector in observations by Equation (2):
(2)
Obtain the expected coefficients matrix for the acquired data by Equation (3):
1
1
(3)
Compute the appropriate equations and as an eigenvalue where
.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
9
Using the starting parameters, produce the necessary components by Equation
(4):
(4)
Attempts to explain the first variation in the data set as much as possible, and
attempts to explain the remaining variance, etc. A few instances of bigger eigenvalues
usually control the rest in the most valuable data sets that are represented in Equation
(5).
(5)
where represents the percentage preserved in the data forms. Since the produced
principal components via PCA extracted features constitute the variability in the data,
they ought to be retained.
3.4. Egyptian vulture optimized adjustable long short-term memory
network (EVO-ALSTM)
Hybrid approach combining the evolutional optimization approach of the
Egyptian Vulture optimized (EVO) with adjustable LSTM (ALSTM) algorithms. The
optimization mimics the foraging characteristic of Egyptian vultures to fine-tune
hyperparameters, including learning rate, amount of hidden units, and dropout rate
within the ALSTM, which is designed to regulate its memory and neglect gates
dynamically. This hybridization approach improves the ALSTM’s capacity to evolve
to varying temporal dependencies in movement classification responsibilities. The
integration ensures faster convergence and better motion sequence reputation
performance by leveraging each efficient optimization and flexible memory
modification in the LSTM. It achieves improved accuracy and efficiency in movement
classification compared to standard LSTM networks.
3.4.1. ALSTM
ALSTM is a variant of the LSTM network designed for motion classification
tasks, wherein the model dynamically adjusts its internal memory gates based on
various temporal dependencies in movement sequences. This adaptability allows the
ALSTM to selectively consider or forget information at special time steps, enhancing
in classifying complex and time-sensitive movement patterns. There are four layers in
the ALSTM prediction model: input, hidden, output, and EVO. The prediction model
is optimized by the usage of EVO. Based on the RNN design approach and
consider the properties. denotes the entire collection, it is
separated into two subsets: a training data set called and
a test data set called. Each of those samplings is
represented by the symbol
′′′′, which includes n
attributes . Utilizing the Z-score
normalization technique (mean 0, variance 1), standardize the training set as expressed
in Equation (6):
Molecular & Cellular Biomechanics 2024, 21(4), 468.
10
′
(6)
In the above instance Ws’ is the standardized amount for the property at
time′ and data are split to adjust to the hidden layer’s input
properties. Equations (7)–(9) denote the data set separation appears as follows when
the separation length is set to:
(7)
′
′
′
(8)
′
′
′
′
(9)
The amount of
′ at a given time is entered as an algorithm in the hidden layer,
which is composed of identical LSTM units interconnected at the times before and
after. The hidden layer’s output in Equation (10) is as follows:
′
(10)
where the current state and output of the preceding LSTM unit have the following
values: and. is the LSTM unit’s forward computation. Equation
(11) represents the final output of the LSTM network after it has passed through a fully
connected layer, which is the ALSTM’s output.
(11)
The training method uses the mean square error, which is defined in Equation
(12), as the loss function.
2
1
(12)
where and stand for the isomorphic LSTM unit and the total amount of input
samples, correspondingly. The numerical value of the LSTM model’s training output
is indicated by. The sample’s real value is denoted by. The network algorithm’s
layer counts, training intervals, and hidden layer neuron count are all simultaneously
adjusted to the optimization.
3.4.2. EVO
EVO was a meta-heuristic method initially developed to solve complicated
arrangement problems. It is aroused by the Egyptian vulture’s behavior to obtain a
solution. This avian creature’s cunning behavior is converted into an algorithm that
can handle challenging optimization issues. The system has been modified to
incorporate the detailed EVOA processes. It optimizes the selection of key features,
improving the accuracy and efficiency of ALSTM in recognizing complicated human
motions. By integrating this bio-inspired optimization technique, the ALSTM can
better capture temporal dependencies and diffused variations in movement
information. This method leads to greater specific and dependable motion types,
Molecular & Cellular Biomechanics 2024, 21(4), 468.
11
especially in dynamic environments. Figure 3 represents the process of EVO,
algorithm 1 represents the EVO-ALSTM algorithm, and the following are the steps in
the procedure:
Figure 3. Process of EVO.
Algorithm 1 EVO-ALSTM algorithm
1: Step 1: def initialize_population():
2: population = []
3: for i in range(pop_size):
4: individual = random_hyperparameters()
5: population.append(individual)
6: return population
7: Step 2: def fitness (individual, train_data, val_data):
8: lstm_model = build_lstm(individual)
9: lstm_model.train(train_data)
10: performance = evaluate_model (lstm_model, val_data)
11: return performance
12: Step 3: def egyptian_vulture_optimization(population, train_data, val_data):
13: for generation in range (max_generations):
14: for vulture in population:
15: fitness_score = fitness (vulture, train_data, val_data)
16: update_best_solution(fitness_score)
17: population = evolve_population(population)
18: return best_solution
19: Step 4: def adjustable_lstm (input_data, hyperparameters):
20: lstm_layer = LSTM (hyperparameters[‘hidden_units’], return_sequences=True)
21: dropout_layer = Dropout(hyperparameters[‘dropout’])
22: adjusted_output = adjust_memory_for_sequence(lstm_layer, input_data)
23: return dropout_layer(adjusted_output)
24: Step 5: def evo_alstm_motion_classification (train_data, val_data):
25: population = initialize_population ()
26: best_hyperparameters = egyptian_vulture_optimization (population, train_data, val_data)
27: lstm_model = build_lstm (best_hyperparameters)
28: Step 6: hyperparameters
29: lstm_model.train (train_data)
30: motion_classification_results = lstm_model.classify (test_data)
31: return motion_classification_results
Molecular & Cellular Biomechanics 2024, 21(4), 468.
12
Step 1: The initialization of the solution set of strings contains changeable
representations of the parameters. One possible solution state is represented by a string
with a set of parameters.
Step 2: Conditions are verified, limitations are superimposed, and representative
variables are refined.
Step 3: Stones are thrown at predetermined or random locations.
Step 4: Either a portion of the string or the complete one is picked for the Rolling
of the Twigs performance.
Step 5: The strategy of changing the angle is used to reverse a specific portion of
the solution.
Step 6: Fitness is assessed.
Step 7: Checking the stopping criterion is necessary.
4. Result and discussion
In this section, the confusion matrix for classification is calculated, the
performance of the proposed method based on the factors including walking, jumping,
arm waving, and sports action is evaluated, and the effectiveness of the proposed
method EVO-ALSTM with the conventional technique ALSTM is compared based on
the metrics (accuracy, precision, recall, and F1-score).
4.1. Experimental setup
An Intel i7-7500U CPU running at 2.70GHz with 8 GB of RAM and Mat lab
R2014a was used to simulate the proposed method.
4.2. Confusion matrix
Figure 4. Output of confusion matrix with 30 participants.
To evaluate the performance of a classification method, it compares the
prediction classification with the actual results by organizing outcomes into four
classes, including true positive (TP), false positive (FP), true negative (TN), and false
Molecular & Cellular Biomechanics 2024, 21(4), 468.
13
negative (FN). Figure 4 represents the output of the confusion matrix with 30
participants. In Figure 4, the diagonal values represent accurate detections, with
excessive accuracy for walking (92), jumping (75), arm waving (96), and sports action
(75). Misclassifications are seen, which include walking expected as sports action (7
instances) and jumping misclassified as sports action (7 instances). Darker cells imply
better accuracy, at the same time as lighter cells reflect fewer correct classifications.
Overall, the model successfully classifies all instances without errors.
4.3. Motion detection accuracy
The motion detection accuracy that estimates how well the system features the
ability to cope with the concerns of identification of movement within an environment
or in a video frame movement tracking. It is expressed as the number of correctly
identified motion events over the number of existed motion events. The accuracy level
of various human activities classified using EVO-ALSTM was evaluated, which is
collected using sensors. Table 1 represents the motion detection accuracy of human
activity.
Table 1. Motion detection accuracy of human activity.
Human activities
Motion detection accuracy (%)
Walking
94.7%
Jumping
92.3%
Arm waving
95%
Sports action
93.4%
Figure 5 shows the wearable sensors and a biomimetic vision system used to
record various human activities indicating reliable movement detection accuracy.
Among the different activities, 94.7 % accuracy was recorded in walking while arm
waving was slightly higher at 95%. Complementing this, jumping and sports-related
activities recorded an accuracy of 92.3 % and 93.4 % respectively. This underlines the
efficient performance of the system in the execution of any other dynamic activities.
According to the findings, the proposed method achieved a high accuracy level in
classifying arm-waving movement.
Figure 5. Motion detection accuracy of human activity.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
14
4.4. Comparison
4.4.1. Accuracy
The model’s accuracy is calculated by dividing the total number of true positive
(TP) and true negative (TN) predictions by the total number of forecasts. This gives
an overall performance level of the model with consideration of all the classes. It
consists of the proportion of accurately identified cases to all occurrences of human
activity. Table 2 and Figure 6 display the performance of accuracy.
Table 2. Values of four metrics.
Methods
Accuracy (%)
Precision (%)
Recall (%)
F1-score (%)
ALSTM
88%
83%
86%
82%
EVO-ALSTM
95%
93%
90%
91%
Figure 6. Performance of accuracy.
Figure 6 compares the accuracy of two methods: ALSTM and EVO-ALSTM
(Proposed). EVO-ALSTM demonstrates improved accuracy achieved 95%, while
ALSTM shows slightly lower performance around 88%. This indicates that the
proposed EVO-ALSTM method outperforms the standard ALSTM in terms of
accuracy.
4.4.2. Precision
Precision focuses on the number of true positive predictions when compared to
the total amount of accurate forecasts that the model produced. This indicates that
based on the model’s overall number of accurate predictions during the period under
consideration, how many of them are real predicted positive cases and how better the
model in making non-positive predictions. It measures how the model accurately
classifies a motion. It indicates how many positive instances are positive. Figure 7
and Table 2 show the precision evaluation.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
15
Figure 7. Precision performances.
Figure 7 depicts an assessment of the precision of ALSTM and the EVO-
ALSTM model. The precision of the EVO-ALSTM technique is remarkably better 93%
which confirms its capability to locate pertinent instances with minimal errors. The
ALSTM achieves an accuracy level of about 83%, which infers a relatively higher
false positive ratio.
4.4.3. Recall
Sensitivity also referred to as recall is the measure of the ratio of the number of
true positives as compared to the number of positive cases in the population. This
shows the capability of the model to recognize all the relevant elements of a certain
class. Recall measures the classifier’s capacity to find every relevant occurrence of a
given motion by calculating the proportion of the true positive to the actual positive.
Table 2 and the recall performance is shown in Figure 8.
Figure 8. Performance of recall.
In Figure 8, the recall percentage in the two methods ALSTM and EVO-ALSTM
is presented with the proposed EVO-ALSTM with very high recall 90% proving its
effectiveness in instances recovery task. The recall performance of ALSTM is less
Molecular & Cellular Biomechanics 2024, 21(4), 468.
16
than that, at about 86%, suggesting that this method loses more relevant information
when compared to EVO-ALSTM.
4.4.4. F1-score
The F1 score is a metric that takes into account both precision and recall by
averaging their numerical scores. It is especially positive in cases of irregular data
distribution where both false positive and false negative are important. It is the balance
between precision and recall, ensuring that the model performs well in both classifying
and accurately assigning motion. Table 2 and the analysis of the F1-score are shown
in Figure 9.
Figure 9. Evaluation of F1-score.
Figure 9 shows the F1 score for both the ALSTM and EVO-ALSTM approaches.
The efficiency of the suggested EVO-ALSTM in various recovery tasks is
demonstrated by its extremely high F1 score of 91%. However, ALSTM’s F1 score
performance is lower than that, at around 82%, indicating that this approach loses more
pertinent data than EVO-ALSTM.
4.5. Discussion
The standard method EVO, while effective for hyperparameters optimization,
could struggle with local minima and convergence speed, particularly in high
dimensional search spaces. Additionally, it is computationally expensive, as it requires
multiple evaluations of the fitness function, a process that can be time-consuming in
large data sets. Moreover, EVO lacks the dynamic adjustment that is necessary for
real-time motion classification because it is designed to deal only with parameters
rather than model structures. Compared to it, ALSTM has a flexible memory
mechanism with hyperparameter sensitivity, and the model cannot adapt to the variety
of motion patterns correctly without tuning hyperparameters. Further, ALSTM can
end up in an overfitting situation, particularly when few data are used in training and
its efficiency tends to drop sharply when input information is noisy. The EVO-
ALSTM models are assisted by the evolutionary algorithms for adaptive optimization,
which improves the ability of the model to optimize and escape out of local minima
reducing the overall convergence times. This allows for improved generalization and
accuracy when dealing with complex and nonlinear data distributions. Furthermore,
Molecular & Cellular Biomechanics 2024, 21(4), 468.
17
due to its flexible structure, the architecture can be customized to fit the specific
attributes of the dataset, thereby enhancing performance. The proposed EVO-ALSTM
method addressed these limitations by integrating the strengths of both techniques: it
leverages the efficient exploration that EVO affords to dynamically configure the
ALSTM architecture, ensuring that the model not only learns effectively but also
manages memory in real-time. This hybridization results in enhanced accuracy and
robustness in motion classification tasks, effectively managing the weakness of the
individual methods and enhancing overall performance.
5. Conclusion
In this paper, an Egyptian Vulture optimized Adjustable Long Short-Term
Memory Network (EVO-ALSTM) was introduced for motion classification. The
dataset was gathered from 30 participants. The wearable sensors, like accelerometers
and gyroscopes, were used to collect human activity, including walking, jumping, arm
waving, and sports actions, from the participants. Preprocessed the data by using Z
score normalization and extracted the complex features by using PCA. The proposed
EVO-ALSTM method was used as a classification to identify the motions. As a result,
the four human activities measured for the motion detection accuracy of the proposed
method showed that arm waving (95%) has a high detection motion accuracy level.
The proposed EVO-ALSTM method was compared with the standard method
(ALSTM) based on the metrics, including accuracy (95%), precision (93%), recall
(90%), and F1-score (91%). According to the findings, the proposed method has
superior performance than other methods to classify human activity and it helps to
enhance the animation design.
Limitation and future scope
The framework’s dependence on wearable sensors may restrict the range of
motion statistics, and the computational complexity of the EVO-ALSTM approach
will be resource-intensive for real-time processing. Additionally, generalization to
various person environments and motions may require further optimization. This
investigation opens avenues for similar studies into integrating biomimetic
imaginative and visual structures with greater superior device mastering models for
even greater accuracy in motion popularity. Future research may want to explore
expanding the variety of human sports recorded and integrating this technology into
various packages like gaming, digital learning, and interactive storytelling. Enhancing
actual time feedback capabilities in VR and AR environments could also increase
immersion.
Ethical approval: Not applicable.
Conflict of interest: The author declares no conflict of interest.
References
1. Wee, C., Yap, K.M. and Lim, W.N., 2021. Haptic interfaces for virtual reality: Challenges and research directions. IEEE
Access, 9, pp.112145-112162.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
18
2. Wang, G., Zheng, C., Fu, Y., Zhu, K., Lai, F., Zhang, L., Li, M., Wu, X., Ren, M., Zheng, Y. and Lian, B., 2024, July.
KiPneu: Designing a Constructive Pneumatic Platform for Biomimicry Learning in STEAM Education. In Proceedings of
the 2024 ACM Designing Interactive Systems Conference (pp. 441-458).
3. Han, L., Afzal, N., Wang, Z., Wang, Z., Jin, T., Guo, S., Gong, H. and Wang, D., 2024. Ambient haptics: bilateral interaction
among humans, machines, and virtual/real environments in the pervasive computing era. CCF Transactions on Pervasive
Computing and Interaction, pp.1-33.
4. Sauda, E., Karduni, A. and Lanclos, D., 2024. Architecture in the Age of Human-Computer Interaction. Taylor & Francis.
5. CHEN, Y., WU, X., ZHANG, J., LIU, Y. and LI, H., 2024. A Type of Human-Computer Collaborative.
6. Coban, M. and Coştu, B., 2023. Integration of biomimicry into science education: Biomimicry teaching approach. Journal of
Biological Education, 57(1), pp.145-169.
7. Soliman, M.E. and Bo, S., 2023. An innovative multifunctional biomimetic adaptive building envelope based on a novel
integrated methodology of merging biological mechanisms. Journal of Building Engineering, 76, p.106995.
8. Jović, B.S. and Mitić, A.D., 2020. Exploration of nature-based biomimetic approach in landscape architectural design: a
parametric study of candelabra model design. Visual Computing for Industry, Biomedicine, and Art, 3, pp.1-11.
9. Li, S., 2021. Ancient Architecture Animation Design Method of 3D Technology and Its Application. In Journal of Physics:
Conference Series (Vol. 2037, No. 1).
10. Pezent, E., Macklin, A., Yau, J.M., Colonnese, N. and O’Malley, M.K., 2023. Multisensory Pseudo‐Haptics for Rendering
Manual Interactions with Virtual Objects. Advanced Intelligent Systems, 5(5), p.2200303.
11. Gallerani, M., Vazzoler, G., De Novi, G., Razzoli, R., Berselli, G. and Ottensmeyer, M.P., 2023. Integrated design and
prototyping of a robotic eye system for ocular and craniofacial trauma simulators. International Journal on Interactive Design
and Manufacturing (IJIDeM), 17(6), pp.3103-3116.
12. Shreyas, D.G., Raksha, S. and Prasad, B.G., 2020. Implementation of an anomalous human activity recognition system. SN
Computer Science, 1(3), p.168.
13. Mahbub, U. and Ahad, M.A.R., 2022. Advances in human action, activity, and gesture recognition. Pattern Recognition
Letters, 155, pp.186-190.
14. Minh Trieu, N. and Thinh, N.T., 2024. Advanced Design and Implementation of a Biomimetic Humanoid Robotic Head
Based on Vietnamese Anthropometry. Biomimetics, 9(9), p.554.
15. McMahon, M. and Erolin, C., 2024. Biomimicry–medical design concepts inspired by nature. Journal of Visual
Communication in Medicine, 47(1), pp.27-38.
16. Dai, F. and Li, Z., 2024. Research on 2D Animation Simulation Based on Artificial Intelligence and Biomechanical
Modeling. EAI Endorsed Transactions on Pervasive Health and Technology, 10.
17. Li, Y., Zhao, M., Yan, Y., He, L., Wang, Y., Xiong, Z., Wang, S., Bai, Y., Sun, F., Lu, Q. and Wang, Y., 2022.
Multifunctional biomimetic tactile system via a stick-slip sensing strategy for human-machine interactions. NPJ Flexible
electronics, 6(1), p.46.
18. Reddy, E.V., Manideep, K. and Agarwal, O., 2024. Machine Learning-Based Human Movements Mimicking System for
Animation and Virtual Reality. Asian Journal of Research in Computer Science, 17(7), pp.84-94.
19. Gao, Q., 2022. Design and Implementation of a 3D Animation Data Processing Development Platform Based on Artificial
Intelligence. Computational Intelligence and Neuroscience, 2022(1), p.1518331.
20. Vignesh, T., 2021, March. Pipeline for Development of 3-dimensional motion animation using 2-dimensional video. In 2021
7th International Conference on Advanced Computing and Communication Systems (ICACCS) (Vol. 1, pp. 991-995). IEEE.
21. Kong, D., Yang, G., Pang, G., Ye, Z., Lv, H., Yu, Z., Wang, F., Wang, X.V., Xu, K. and Yang, H., 2022. Bioinspired Co‐
Design of Tactile Sensor and Deep Learning Algorithm for Human–human-robot interaction. Advanced Intelligent Systems,
4(6), p.2200050.
22. Kim, S.J., Lee, Y.J. and Park, G.M., 2021. Real-Time Joint Animation Production and Expression System using Deep
Learning Model and Kinect Camera. Journal of Broadcast Engineering, 26(3), pp.269-282.
23. Fu, Q., Fu, J., Zhang, S., Li, X., Guo, J. and Guo, S., 2021. Design of intelligent human-computer interaction system for hard
of hearing and non-disabled people. IEEE Sensors Journal, 21(20), pp.23471-23479.
24. Qin, X., Xia, X., Ge, Z., Liu, Y. and Yue, P., 2024. The Design and Control of a Biomimetic Binocular Cooperative
Perception System Inspired by the Eye Gaze Mechanism. Biomimetics, 9(2), p.69.
Molecular & Cellular Biomechanics 2024, 21(4), 468.
19
25. Lu, W., 2024. Learning-Based, Muscle-Actuated Biomechanical Human Animation: Bipedal Locomotion Control and Facial
Expression Transfer (Doctoral dissertation, UCLA).
26. Zeng, X.S., Dwarakanath, S., Lu, W., Nakada, M. and Terzopoulos, D., 2021. Neuromuscular Control of the Face-Head-
Neck Biomechanical Complex with Learning-Based Expression Transfer from Images and Videos. In Advances in Visual
Computing: 16th International Symposium, ISVC 2021, Virtual Event, October 4-6, 2021, Proceedings, Part I (pp. 116-127).
Springer International Publishing.
27. Kumarapu, L. and Mukherjee, P., 2021. Animepose: Multi-person 3d pose estimation and animation. Pattern Recognition
Letters, 147, pp.16-24.
28. Jones, M., Byun, C., Johnson, N. and Seppi, K., 2023. Understanding the Roles of Video and Sensor Data in the Annotation
of Human Activities. International Journal of Human-Computer Interaction, 39(18), pp.3634-3648.
29. Yadav, R.K., Arockiam, D. and Bhaskar Semwal, V., 2024. Motion Signal-based Recognition of Human Activity from
Video Stream Dataset Using Deep Learning Approach. Recent Advances in Computer Science and Communications
(Formerly: Recent Patents on Computer Science), 17(3), pp.77-91.
