ChapterPDF Available

Real-Time Embedded Person Detection and Tracking for Shopping Behaviour Analysis


Abstract and Figures

Shopping behaviour analysis through counting and tracking of people in shop-like environments offers valuable information for store operators and provides key insights in the stores layout (e.g. frequently visited spots). Instead of using extra staff for this, automated on-premise solutions are preferred. These automated systems should be cost-effective, preferably on lightweight embedded hardware, work in very challenging situations (e.g. handling occlusions) and preferably work real-time. We solve this challenge by implementing a real-time TensorRT optimized YOLOv3-based pedestrian detector, on a Jetson TX2 hardware platform. By combining the detector with a sparse optical flow tracker we assign a unique ID to each customer and tackle the problem of loosing partially occluded customers. Our detector-tracker based solution achieves an average precision of 81.59% at a processing speed of 10 FPS. Besides valuable statistics, heat maps of frequently visited spots are extracted and used as an overlay on the video stream.
Content may be subject to copyright.
Real-time Embedded Person Detection and
Tracking for Shopping Behaviour Analysis
R. Schrijvers1,2, S. Puttemans1, T. Callemein1, and T. Goedem´e1
1EAVISE, KU Leuven, Jan De Nayerlaan 5, 2860 Sint-Katelijne-Waver, Belgium
2PXL Smart-ICT, Hogeschool PXL, Elfde Liniestraat 24, 35000 Hasselt, Belgium
{steven.puttemans, toon.goedeme, timothy.callemein}
Abstract. Shopping behaviour analysis through counting and tracking
of people in shop-like environments offers valuable information for
store operators and provides key insights in the stores layout (e.g.
frequently visited spots). Instead of using extra staff for this, automated
on-premise solutions are preferred. These automated systems should
be cost-effective, preferably on lightweight embedded hardware, work
in very challenging situations (e.g. handling occlusions) and preferably
work real-time. We solve this challenge by implementing a real-time
TensorRT optimized YOLOv3-based pedestrian detector, on a Jetson
TX2 hardware platform. By combining the detector with a sparse optical
flow tracker we assign a unique ID to each customer and tackle the
problem of loosing partially occluded customers. Our detector-tracker
based solution achieves an average precision of 81.59% at a processing
speed of 10 FPS. Besides valuable statistics, heat maps of frequently
visited spots are extracted and used as an overlay on the video stream.
Keywords: Person Detection ·Person Tracking ·Embedded ·Real-time
1 Introduction
Mapping the flow and deriving statistics (e.g. the amount of visitors or the time
spent in as store) of people visiting shop-like environments, holds high value for
store operators. To this day, people counting in retail environments is often being
accomplished by using cross-line detection systems [24] or algorithms counting
people through virtual gates [32].
To accurately count visitors, one can place a computer in the network of the
store with access to already available security cameras, deploying software that
detects and tracks people in the store automatically and stores the results on
a central storage system (e.g. an in-store server or the cloud). In order to run
these software solutions, one needs expensive and bulky dedicated computing
hardware, frequently covered by a desktop GPU (e.g. NVIDIA RTX 2080). On
the other hand, the recent availability of lightweight and affordable embedded
GPU solutions, like the NVIDIA Jetson TX2, can be a valid alternative. This
is the main motivator to build an embedded and cost-effective people detection
2 R. Schrijvers; S. Puttemans; et al.
Fig. 1. Examples of (left) the viewpoint from our camera setup in the store and (right)
the generated heat map for the store owner using the combined detector-tracker unit.
and tracking solution. An example of the camera viewpoint from the designed
setup can be seen in the left part of Figure 1.
The remainder of this paper is organized as follows. Section 2 discusses related
work on person detection and tracking, along with available embedded hardware
solutions. Section 3 provides details about the proposed implementation, while
experiments and results are discussed in section 4. Finally, section 5 summarizes
this work and discusses useful future research directions.
2 Related Work
The related work section is subdivided in three subsections, each focusing on
a specific subtopic within this manuscript. Subsection 2.1 starts by discussing
literature on person detection, where subsection 2.2 continues on specific
optimizations in this technology for embedded hardware. Finally subsection 2.3
focuses on the person tracking part.
2.1 Person Detection
Robust person detection solutions have been heavily studied in literature.
Originally person detection made use of handcrafted features, combined with
machine learning to generate an abstract representation of the person [6, 9, 47,
15, 12, 13]. While these approaches showed promising results, their top-accuracy
and flexibility make that these algorithms are not suited for the very challenging
conditions in which they will be deployed in this application. Dynamic
backgrounds, illumination changes and different store lay-outs are only some
of the reasons of a high rate of false positive detections. While re-training the
detectors for every store layout would theoretically solve the issue, this isn’t a
cost-effective solution for companies.
Convolutional neural networks for detecting persons in images offer
potentially more robust solutions [20, 31, 38, 39, 17]. By automatically selecting
the most discriminate feature set based on a very large set of training
data, they quickly pushed the traditional approaches into the background. In
Real-time Embedded Person Detection and Tracking 3
literature these deep learning based approaches are subdivided in two categories.
Single-shot approaches [31, 38–40] solve the detection task by classifying and
proposing bounding boxes in a single feed-forward step through the network.
Multi-stage approaches [17, 19, 16, 41] include separate networks that first
generate region proposals before classifying the objects inside the proposals.
A more recent and promising multi-stage approach is Trident-Net [30], where
scale-specific feature maps are built into the network. Single-stage approaches
tend to have a more compact and faster architecture making them preferred in
lightweight embedded hardware solutions.
2.2 Optimizing for embedded hardware
Solving the task of person detection on embedded platforms has been an active
research topic in recent years. A common approach is to enable deep learning on
embedded platforms by studying more compact models such as Tiny-YOLOv2
[39]. While these compacter models are able to run at decent speeds on embedded
hardware, they tend to lose some percentage in accuracy compared to their full
counterparts (e.g. YOLOv2). Besides going compact, several approaches [43, 49,
34] optimize the architecture further by looking at the indirect computation
complexity and address efficient memory access and platform characteristics.
More recent embedded object detection algorithms introduce optimized filter
solutions, like depth-wise separable convolutions [22] and inverted residuals [42],
to further optimize the performance of these embedded solutions.
With the uprising of FPGA chips for deep learning, several architectures
like [23] try to reduce the number of parameters of the models even further,
having a 50x reduction in parameters compared to classic models like [28].
These specifically designed hardware chips also allow for fixed-point 16-bit
optimizations through OpenCL [48]. While being very promising, these FPGA
systems still lack flexibility. They are in most cases designed for a very specific
case, and are thus no cost-effective solution for our problem.
While many of these embedded object detectors are explicitly shaped for
running on embedded hardware, detection accuracies are still lower compared
to their traditional full blown CNN counterparts (eg. YOLOv3). Taking into
account that our solution should perform person detection at a minimal
processing speed of 10 FPS and a minimal accuracy of 80%, we opt for integrating
an optimized embedded implementation of the YOLOv3 object detector [46] in
our pipeline. The architecture smartly combines several optimizations based on
mobile convolutions in PyTorch and TensorRT compilations.
The introduction of the NVIDIA Jetson embedded GPU enabled balancing
of local-processing, low power consumption and throughput in an efficient way.
[35] discusses several implementations of deep learning models on the Jetson
platform while considering a range of applications, e.g. autonomous driving or
traffic surveillance. The work focuses on obtaining low latency, to make detectors
useful for providing valuable real-time feedback. The advantages in both FPGA
and embedded GPU systems are the fact that they are substantially more
space-efficient and less power-consuming than desktop GPUs.
4 R. Schrijvers; S. Puttemans; et al.
2.3 Person Tracking
Tracking objects in videos has been studied across many fields. Whenever
quasi-static environments occur, motion detection through robust background
subtraction is used to identify moving objects in images [11, 8, 29, 27]. More
challenging cases involve dynamic environments, e.g. tracking objects in
autonomous driving applications or drones. [7] solves the tracking task
using detection results of a lightweight object detection algorithm combined
with euclidean distance equations, GPS-locations and data from an inertial
measurement unit. Because research is moving more and more towards highly
accurate CNN-based object detectors, a new sort of tracking is introduced, called
tracking-by-detection. By calculating an intersection over union (IoU) between
the detections of two consecutive frames and applying a threshold, one can decide
if we are dealing with the same object.
Where tracking-by-detection works well in many situations, there are also
some downsides with these approaches. When people are missed in several
subsequent frames by the initiating detector, the person gets lost during
tracking. However, in our application of real-time costumer analysis, we want
to keep a unique ID for each detected person, and thus need to fill in
this gap automatically. [2] solves this issue by integrating a SVM-classifier
with an optical-flow-based tracker. With the uprising of CNN-based detectors,
CNN-based object trackers have also been proposed. Limb-based detection with
tracking-by-detection is proposed in [1], while [4] explains an approach using IoU
information from multiple object detectors. In addition, [36] combines location
information and similarity measures to perform tracking-by-detection.
Feature-based tracking algorithms are a valid alternative. Sparse optical
flow calculates the flow of objects based on a sparse feature set [33, 45] and
have been successfully deployed for person tracking [26, 44, 11]. Dense optical
flow [14] calculates the flow for every point in the detection and is thus more
computationally complex. [37] propose an object tracker based on weakly aligned
multiple instance local features and an appearance learning algorithm.
A final range of trackers use online learning, where a tracker learns the
representation of the object on the fly through a classifier when initialized with a
bounding box [18, 3]. [21] learns more efficiently from these overlapping patches
by using kernelized correlation filters, while [25] enables tracking failure detection
by tracking objects both forward and backward in time, measuring and qualifying
the discrepancies between the two trajectories.
[10] gives a clear general overview of state-of-the-art trackers and their
accuracy on a public dataset. For this work, we integrated three object trackers:
sparse optical flow, kernelized correlation filters and median flow.
This paper proposes a complete off-the-shelf solution, taking into account
the fact that we work on a compact embedded system with limited power
consumption, achieve real-time performance (e.g. minimally processing at 10
FPS) and obtain acceptable accuracies (e.g. over 75%) in a single setup. While
many of these sub-tasks have been discussed already in literature, to our best
knowledge, there isn’t a single publication that proposes such end-to-end solution
Real-time Embedded Person Detection and Tracking 5
Fig. 2. Pipeline overview with the proposed batch processing approach, passing frames
from the image buffer to the detection or tracking unit, based on the batch iterator.
for in-store customer behaviour analysis. On top of that we are combining all
parts of the pipeline in a batch system that consumes the resources (both CPU
and GPU) of the host system as efficient as possible. This is the biggest novelty
this work introduces and will be discussed in further detail in subsection 3.1.
3 Methodology
The goal of this paper is to map the flow of costumers in shop-like environments,
with a special focus on detecting and tracking individual costumers using a
unique ID. Preferably the system should run minimally at a processing speed
of 10 frames per second on top of a compact, easy-to-use, embedded platform
(e.g. Jetson TX2). Therefore, we propose a multi-threaded YOLOv3 TensorRT
optimized person detector combined with a fast lightweight object tracker.
Figure 2 gives an overview on how our pipeline is implemented on the
embedded Jetson TX2 platform. All sub-parts will be discussed in detail in
the following subsections.
3.1 Person detection and batch processing
In applications that require real-time performance, one can either choose to
maximize throughput or minimize latency. Throughput focuses on the number
of images that are processed in a given time slot, while latency focuses on the
time needed to process a single image. Given the goal of analyzing the flow of
costumers in a given security camera stream, and in the future even in multiple
camera streams, throughput is in our case more important than latency. We
maximize throughput by performing a batch processing approach that takes
advantage parallel GPU processing. As seen in Figure 2, an image buffer is
used to collect all incoming images from the camera stream, at a 1280x720
resolution, which are downsampled to a 512×288 input resolution before moving
to the detector and the tracker unit. We could sent these images directly to our
optimized YOLOv3 implementation [46], but since it is only capable of processing
6 R. Schrijvers; S. Puttemans; et al.
data at a maximum speed of 8FPS at this resolution, we skip several frames when
collecting a batch that needs to be processed by the YOLOv3 detector.
Take for example a image buffer of 24 images. From those 24 images, only
8 images are selected as a batch that gets passed to the detector unit. The
remaining 16 images are sent to the tracker unit, which waits for the processed
batch of the detector, so that it can track each detected person throughout the
24 frames series, based on the detections in those 8 reference frames. In order
to improve processing speed, both detector and tracker unit are implemented in
separate threads, to reduce the processing time delays again as much as possible.
3.2 Person tracking
Given our hardware setup of a Jetson TX2 embedded platform, we need to
carefully consider our hardware resources. Since we’re using the on-board GPU
to efficiently processes batches of 8 images with the object detector unit, the
decision was made to implement a CPU-based tracker to divide the workload as
good as possible between the different processing units. We augment the detector
unit with a lightweight CPU-based Lukas Kanade Sparse Optical Flow tracker
(based on the OpenCV4.1 implementation [5]). This allows us to run the tracker
on the CPU, while simultaneously running the GPU-focused YOLOv3 object
detector. To achieve the minimally required frame rate of 10 FPS, we downscale
the original input frames to a resolution of 512×288 pixels.
In between detection frames, two extra frames remain which are only
processed by the tracker unit, and thus so far have no knowledge of the location
of the detected objects. For each detection inside a detection frame, key feature
points are calculated, generating a sparse representation of that detection. Those
points are passed into an optical flow mechanism that is used to generate
predictions of the locations of these keypoints in the in-between tracker frames.
Finally the detections in the next detection frame are used to apply a mapping
between the predictions and the detections, possible proposing a slight correction
of the bounding box location. Figure 3 illustrates how this detector-tracker
combination is working on a sample from the security camera stream. The tracker
only produces predictions in the next frame for detections with a detection
Fig. 3. The detector-tracker pipeline, using tracking-by-detection with interleaved
tracker proposals: (solid red) detections (interleaved green) tracker proposals (solid
green) tracker proposal accepted as detection in case of a missing detection.
Real-time Embedded Person Detection and Tracking 7
probability higher than 10%. This allows us to carefully choose the working
conditions of the detector and avoid as much false negative detections (persons
getting missed) as possible. Detections with a lower confidence will simply be
ignored by the tracker and immediately be removed from the tracker memory.
Applying this approach also introduces issues. When customers move outside
the field-of-view of the camera stream, their tracking information is kept in
memory by the tracker, clogging up the tracker memory and resulting in locally
dangling tracks. In order to avoid these dangling tracks, after 5 consecutive
detection misses, we force the tracker to remove the tracking information and
forget the track altogether, which will end the tracking of that object.
Besides the sparse optical flow approach, two more CPU-based tracking
implementations (Kernelized Correlation Filers and Median Flow) are tested to
compare tracking robustness in shop-like environments. Both are initialized by
the detections of the neural network, just like the sparse optical flow approach.
For both trackers, the same input resolution of 512x288 pixels has been used.
3.3 Tracker memory
Since the selected tracker implementations only takes the information between
two consecutive frames into account, we are risking losing a person with a specific
ID from tracker memory (e.g. due to exceeding the threshold of missed frames),
that can in a later stage, be picked up again by the detector. In order to avoid a
new ID being assigned, we introduce a method to keep and match these lost
tracks in memory. In case a track is deleted from memory, it first changes
its status to lost. A lost track is remembered for maximally 5 seconds, and
is recovered whenever a new detection appears in a location close to the last
known location of a deleted track. A match between a lost track and a newly
generated track is made by meeting one of the following conditions: the new
location lies within a radius of 200 pixels around the last known location, or the
new location lies in the quadrant of which the gradient of the lost track points
to. In case of multiple new detections in the same quadrant, the detection with
the shortest (perpendicular) distance to the gradient is preferred.
3.4 Heat map extraction
On top of providing statistics, about the exact location and path followed by
customers, users of the system would also like to have some sort of visual
confirmation on the activity in their stores. Therefore we propose a heat map
based system that gives an overview of the store spots that are most frequently
visited within a specific time slot. The heat maps are extracted by mapping
the pixels in the detected and tracked bounding boxes on top of a visual layout
of the store. The used object detector is known for its jittery bounding boxes,
with inconsistent widths and heights. By simply incrementing pixel locations
each time a detection box matches a pixel, we end up with a very jittered and
visually unpleasing heat map, which we would like to avoid. We solve this by
using a weighted increment of the pixels falling within a bounding box. By
8 R. Schrijvers; S. Puttemans; et al.
taking the ratio between the bounding box width and height into account and
by giving the center pixel a maximum weight that degrades towards the borders
of the bounding box, we end op with a increment value that is less influenced
by the jitter, resulting in a visually pleasing heat map. After all frames within
the given time slot have been processed, heat maps are normalized to get a
meaningful color overlay, as seen in Figure 1.
4 Experiments and results
In this section, we evaluate the proposed embedded detector-tracker approach
used for in-store customer behaviour analysis by performing four different
experiments. We start by evaluating the detector unit and decide on the final
deep learned architecture for the detector unit. Next, we add different trackers
to the problem to cope with missing detections, optimize parameters and have
a look at the influence of changing internal parameters. We then integrate our
optimal detector-tracker combination into our batch processing pipeline, to gain
processing speed. Finally we take a shot at generating visual heat maps, giving
confirmation of frequently visited places in the stores.
4.1 Detector and tracker evaluation
A set of 5000 in-store images were collected via the security camera streams
inside the store and annotated with ground truth labels. We evaluate both the
TensorRT optimized YOLOv2 and YOLOv3 implementations on this set by
running them through our NVIDIA Jetson TX2 platform. The left hand side
of Figure 4 shows the difference in obtained average precision (AP). We notice
that the YOLOv3 architecture (AP=84.36%) is performing much better than
the YOLOv2 architecture (AP=64.26%), with a tremendous rise in AP of 20%.
This can be explained by the capability of the YOLOv3 architecture to detect
persons over a wider variety of scales. Giving the security cameras are mounted
to give a single point overview of the store, customers walking further away
Fig. 4. Precision-recall curves of (left) the embedded and optimized YOLOv2 and
YOLOv3 implementation (right) the combined detector-tracker implementations based
on optical flow, median flow and kernelized correlation filters.
Real-time Embedded Person Detection and Tracking 9
from the camera are not being detected by the YOLOv2 architecture. On top of
that, we also notice that the YOLOv3 architecture is better capable of detecting
partially occluded persons.
However, even with the YOLOv3 architecture, we do not yet reach the
optimal solution (AP=100%). After carefully inspecting the missed detections,
we notice the detect still drops persons when they are crossing each other or
when they are move behind counters and thus get partially occluded. To solve
these issues we run our combined detector-tracker units, for which the AP curves
are visible in Figure 4 at the right hand side. We notice that adding the tracker
unit slightly helps the detector improve in AP for all the combinations. Based
on these experiments we decided to stick with the best performing combination,
being YOLOv3 as detector and optical flow as tracker.
One could however argue that the influence of the tracker can also be
increased if predicted tracks are kept in memory longer. Until now a track is
rejected as soon as the detector unit is unable to find a matching detection
in 5 consecutive detection frames. To show the influence of this parameter, we
increased this threshold towards 10 missed detection frames and again ran the
precision-recall evaluation for the optimal combination again, as seen in Figure
5 on the left hand side. This change again provides a small boost to the AP
of the optimal solution towards 85.35%. However, the influence of varying this
parameter even further should be investigated in future research.
However, up till now our experiments were targeted at getting the highest
AP possible with a detector-tracker combination. Since the original TensorRT
YOLOv3 optimized implementation is only able to run the network at an average
speed of 3FPS, adding the tracker reduced this further towards 2FPS. Since we
want to aim for a minimal processing speed of 10 FPS, this did not suffice. As
discussed in subsection 3.1 we applied a batch processing pipeline to overcome
this low processing speed. The obtained AP is seen in the right part of Figure
5. While increasing our processing speed towards 9.8 FPS by simultaneously
reducing the AP towards 81.59%. In order to see if this change might influence
the other detector-tracker combinations, we also evaluated those in the batch
Fig. 5. Influence of (left) increasing the miss threshold for the tracker and (right)
applying our batch processing pipeline to gain processing speed.
10 R. Schrijvers; S. Puttemans; et al.
Optical Flow Median Filtering Kernelized Correlation Filters
9.8 FPS 4.7 FPS 2.1 FPS
Table 1. Comparing average detector-tracker processing speeds over the dataset.
processing setup. Results of the obtained average tracking speeds can be seen in
Table 1, which confirmed that we picked the correct combination before. Median
filtering and kernelized correlation filters perform worse due to the linear scale-up
in relation to the the amount of detected persons, while this is not true for optical
flow. However, we acknowledge that this could depend on the implementation of
the tracker used, and that further research on this is necessary.
4.2 Generating heat maps of frequently visited places
Figure 1 illustrates the resulting heat maps that indicate the most frequently
visited spots in the store. This gives a clear indication to the shop owners which
counters attract customers better than others. Generating heat maps with the
detector unit only tends to generate heat maps with visually only very small
differences. While the tracker helps in obtaining higher processing speeds in
capturing some of the missed detections due to occlusion, it does not really
benefit the heatmap generation part.
5 Conclusion
In this work we propose a solution for mapping the flow of customers in shop-like
environments in real-time based on person detection and tracking, implemented
on an power-efficient and compact embedded platform. In this work we proposed
a novel batch processing approach that efficiently uses the power of both GPU
and CPU simultaneously, to run a state-of-the-art TensorRT optimized YOLOv3
object detector combined with a sparse optical flow object tracker on the
embedded Jetson TX2 platform. We achieve an average precision of 81.59%
at a processing speed of 10 FPS on a dataset of challenging real-life imagery
acquired in a real shop. While this is not yet the optimal solution, given the
challenging conditions (dynamic backgrounds, illumination changes and different
store lay-outs), this numbers are quite impressive. On top of that we provide
visually pleasing heat maps of the store, giving owners valuable insights to
customer behavior. To improve the current results, we are planning to add a
person re-identification pipeline to the solution and keep an eye out to new ways
of optimizing deep learning pipelines for embedded platforms. This paper creates
a proof-of-concept setup that can be further exploited in multiple camera stream
setups, exploring how we can use shared memory buffers and shared detection
networks between different embedded platforms efficiently.
6 Acknowledgements
This work is supported by PixelVision, KU Leuven and Flanders Innovation &
Entrepreneurship (VLAIO) through a Baekelandt scholarship.
Real-time Embedded Person Detection and Tracking 11
1. Andriluka, M., Roth, S., Schiele, B.: People-tracking-by-detection and
people-detection-by-tracking. In: Proceedings of CVPR. pp. 1–8 (2008)
2. Avidan, S.: Support vector tracking. In: Proceedings of CVPR. pp. 1064–1072
3. Babenko, B., Yang, M.H., et al.: Visual tracking with online multiple instance
learning. In: Proceedings of CVPR. pp. 983–990 (2009)
4. Bochinski, E., Eiselein, V., Sikora, T.: High-speed tracking-by-detection without
using image information. In: Proceedings of AVSS. pp. 1–6 (2017)
5. Bradski, G., Kaehler, A.: Learning OpenCV: computer vision with the OpenCV
library. ” O’Reilly Media, Inc.” (2008)
6. Breiman, L.: Random forests. Machine learning pp. 5–32 (2001)
7. Chen, P., Dang, Y., et al.: Real-time object tracking on a drone with multi-inertial
sensing data. Transactions on ITS pp. 131–139 (2017)
8. Choi, J.W., Moon, D., Yoo, J.H.: Robust multi-person tracking for real-time
intelligent video surveillance. Proceedings of ETRI pp. 551–561 (2015)
9. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In:
Proceedings of CVPR. pp. 886–893 (2005)
10. Dendorfer, P., Rezatofighi, H., et al.: CVPR19 tracking and detection challenge:
how crowded can it get? arXiv preprint arXiv:1906.04567 (2019)
11. Denman, S., Chandran, V., et al.: Person tracking using motion detection and
optical flow. In: Proceedings of DSPCS. pp. 1–6 (2005)
12. Doll´ar, P., Appel, R., et al.: Fast feature pyramids for object detection. Proceedings
of TPAMI pp. 1532–1545 (2014)
13. Doll´ar, P., Tu, Z., et al.: Integral channel features. Proceedings of BMVC pp.
91.1–91.11 (2009)
14. Farneb¨ack, G.: Two-frame motion estimation based on polynomial expansion. In:
Scandinavian conference on Image analysis. pp. 363–370 (2003)
15. Felzenszwalb, P.F., McAllester, D.A., et al.: A discriminatively trained, multiscale,
deformable part model. In: Proceedings of CVPR. pp. 1–8 (2008)
16. Girshick, R.: Fast R-CNN. In: Proceedings of ICCV. pp. 1440–1448 (2015)
17. Girshick, R., Donahue, J., et al.: Rich feature hierarchies for accurate object
detection & semantic segmentation. In: Proceedings of CVPR. pp. 580–587 (2014)
18. Grabner, H., Grabner, M., et al.: Real-time tracking via on-line boosting. In:
Proceedings of BMVC. pp. 47–56 (2006)
19. He, K., Gkioxari, G., et al.: Mask R-CNN. In: Proceedings of ICCV. pp. 2961–2969
20. He, K., Zhang, X., et al.: Deep residual learning for image recognition. In:
Proceedings of CVPR. pp. 770–778 (2016)
21. Henriques, J.F., Caseiro, R., et al.: High-speed tracking with kernelized correlation
filters. Proceedings of TPAMI pp. 583–596 (2014)
22. Howard, A.G., Zhu, M., et al.: Mobilenets: efficient convolutional neural networks
for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)
23. Iandola, F.N., Han, S., et al.: SqueezeNet: AlexNet-level accuracy with 50x fewer
parameters and ¡ 0.5 MB model size. arXiv preprint arXiv:1602.07360 (2016)
24. Iguernaissi, R., Merad, D., et al.: People counting based on kinect depth data. In:
Proceedings of ICPRAM. pp. 364–370 (2018)
25. Kalal, Z., Mikolajczyk, K., et al.: Forward-backward error: automatic detection of
tracking failures. In: Proceedings of CVPR. pp. 2756–2759 (2010)
12 R. Schrijvers; S. Puttemans; et al.
26. Kanagamalliga, S., Vasuki, S.: Contour-based object tracking in video scenes
through optical flow and Gabor features. Optik pp. 787–797 (2018)
27. Kowcika, A., Sridhar, S.: A literature study on crowd (people) counting with the
help of surveillance videos. International Journal of Innovative Technology and
Research pp. 2353–2361 (2016)
28. Krizhevsky, A., Sutskever, I., et al.: Imagenet classification with deep convolutional
neural networks. In: Advances of NeurIPS. pp. 1097–1105 (2012)
29. Lefloch, D., Cheikh, F.A., et al.: Real-time people counting system using a single
video camera. In: Real-Time Image Processing (2008)
30. Li, Y., Chen, Y., et al.: Scale-aware trident networks for object detection. arXiv
preprint arXiv:1901.01892 (2019)
31. Liu, W., Anguelov, D., et al.: SSD: single shot multibox detector. In: Proceedings
of ECCV. pp. 21–37 (2016)
32. Liu, X., Tu, P.H., et al.: Detecting and counting people in surveillance applications.
In: Advanced Video and Signal Based Surveillance. pp. 306–311 (2005)
33. Lucas, B.D., Kanade, T., et al.: An iterative image registration technique with an
application to stereo vision. Proceedings of IJCAI pp. 674–679 (1981)
34. Ma, N., Zhang, X., et al.: ShuffleNet v2: practical guidelines for efficient CNN
architecture design. In: Proceedings of ECCV. pp. 116–131 (2018)
35. Mittal, S.: A survey on optimized implementation of deep learning models on the
NVIDIA Jetson platform. Journal of Systems Architecture (2019)
36. Murray, S.: Real-time multiple object tracking - a study on the importance of
speed. arXiv preprint arXiv:1709.03572 (2017)
37. Pernici, F., Del Bimbo, A.: Object tracking by oversampling local features.
Proceedings of TPAMI pp. 2538–2551 (2013)
38. Redmon, J., Divvala, S., et al.: You only look once: unified, real-time object
detection. In: Proceedings of CVPR. pp. 779–788 (2016)
39. Redmon, J., Farhadi, A.: YOLO9000: better, faster, stronger. In: Proceedings of
CVPR. pp. 7263–7271 (2017)
40. Redmon, J., Farhadi, A.: YOLOv3: an incremental improvement. arXiv preprint
arXiv:1804.02767 (2018)
41. Ren, S., He, K., et al.: Faster R-CNN: towards real-time object detection with
region proposal networks. In: Advances of NeurIPS. pp. 91–99 (2015)
42. Sandler, M., Howard, A., et al.: Mobilenetv2: inverted residuals and linear
bottlenecks. In: Proceedings of CVPR. pp. 4510–4520 (2018)
43. Shafiee, M.J., Chywl, B., et al.: Fast YOLO: a fast you only look once system
for real-time embedded object detection in video. arXiv preprint arXiv:1709.05943
44. Shashev, D., et al.: Methods and algorithms for detecting objects in video files. In:
MATEC Web of Conferences. p. 01016 (2018)
45. Shi, J., Tomasi, C.: Good features to track. In: Proceedings of CVPR. pp. 593–600
46. Vandersteegen, M., Vanbeeck, K., et al.: Super accurate low latency object
detection on a surveillance UAV. arXiv preprint:1904.02024 (2019)
47. Viola, P., Jones, M., et al.: Rapid object detection using a boosted cascade of
simple features. Proceedings of CVPR pp. 511–518 (2001)
48. Wai, Y.J., Mohd Yussof, Z., et al.: Fixed point implementation of Tiny-YOLOv2
using OpenCL on FPGA. Proceedings of IJACSA pp. 506–512 (2018)
49. Zhang, X., Zhou, X., et al.: Shufflenet: an extremely efficient convolutional neural
network for mobile devices. In: Proceedings of CVPR. pp. 506–512 (2018)
In recent years, researchers have made great efforts in computer vision task (e.g., object detection) with the widely use of convolutional neural networks (CNNs). However, object detection algorithms based on CNNs suffer from high computation cost even on the high‐performance computers. In addition, with the development of high‐resolution videos, the deployment of object detection algorithms becomes more and more difficult because of the large amount of data, let alone the portable platforms, such as unmanned aerial vehicles (UAVs). In this paper, we research a lightweight network on portable platform for outdoor tiny pedestrian detection. Concretely, we first set up a training dataset manually for lack of tiny pedestrian samples in common datasets. We provide a lightweight network, and then, parallel computing is introduced to make the most of the advantage of GPU. Finally, our method can achieve real‐time performance on Jetson TX2. Experimental results verify that the proposed model has promising performance in tiny pedestrian detection designed for portable GPU platforms.
Full-text available
Design of hardware accelerators for neural network (NN) applications involves walking a tight rope amidst the constraints of low-power, high accuracy and throughput. NVIDIA's Jetson is a promising platform for embedded machine learning which seeks to achieve a balance between the above objectives. In this paper, we provide a survey of works that evaluate and optimize neural network applications on Jetson platform. We review both hardware and algorithmic optimizations performed for running NN algorithms on Jetson and show the real-life applications where these algorithms have been applied. We also review the works that compare Jetson with similar platforms. While the survey focuses on Jetson as an exemplar embedded system, many of the ideas and optimizations will apply just as well to existing and future embedded systems. It is widely believed that the ability to run AI algorithms on low-cost, low-power platforms will be crucial for achieving the "AI for all" vision. This survey seeks to provide a glimpse of the recent progress towards that goal.
Full-text available
Video files are files that store motion pictures and sounds like in real life. In today's world, the need for automated processing of information in video files is increasing. Automated processing of information has a wide range of application including office/home surveillance cameras, traffic control, sports applications, remote object detection, and others. In particular, detection and tracking of object movement in video file plays an important role. This article describes the methods of detecting objects in video files. Today, this problem in the field of computer vision is being studied worldwide.
Deep Convolutional Neural Network (CNN) algorithm has recently gained popularity in many applications such as image classification, video analytic and object detection. Being compute-intensive and memory expensive, CNN-based algorithms are hard to be implemented on the embedded device. Although recent studies have explored the hardware implementation of CNN-based object classification models such as AlexNet and VGG, there is still a rare implementation of CNN-based object detection model on Field Programmable Gate Array (FPGA). Consequently, this study proposes the fixed-point (16-bit) implementation of CNN-based object detection model: Tiny-Yolo-v2 on Cyclone V PCIe Development Kit FPGA board using High-Level-Synthesis (HLS) tool: OpenCL. Considering FPGA resource constraints in term of computational resources, memory bandwidth, and on-chip memory, a data pre-processing approach is proposed to merge the batch normalization into convolution layer. To the best of our knowledge, this is the first implementation of Tiny-Yolo-v2 object detection algorithm on FPGA using Intel FPGA Software Development Kit (SDK) for OpenCL. Finally, the proposed implementation achieves a peak performance of 21 GOPs under 100 MHz working frequency.