2021 IEEE International Conference on Smart Computing (SMARTCOMP)
An Efficient Edge Deep Learning Computer Vision
System to Prevent Sudden Infant Death Syndrome
2021 IEEE International Conference on Smart Computing (SMARTCOMP) | 978-1-6654-1252-0/21/$31.00 ©2021 IEEE | DOI: 10.1109/SMARTCOMP52413.2021.00061
Vivek Bharati
Homestead High School
Cupertino, California, USA
vbharati238@student.fuhsd.org
Abstract—Sudden Infant Death Syndrome (SIDS) causes in-
fants under one year of age to die inexplicably. One of the most
important external factors, also called an “outside stressor,” that
is responsible for Sudden Infant Death Syndrome (SIDS), is the
sleeping position of the baby. According to past research, the
risk of SIDS increases when the baby sleeps facedown on the
stomach. We propose a Convolutional Neural Network (CNN)
based computer-vision system that estimates the sleeping pose
of the baby and alerts caregivers on their mobile phones within
a few seconds of the baby moving to the hazardous facedown
sleeping position. The system has a low computational load and
a low memory footprint. This characteristic allows the system
to be embedded in low power edge devices such as certain
baby monitors. Processing at the edge also alleviates privacy
concerns with regards to sending images into the network. We
experimented with various numbers of convolutional processing
units and dense layers as well as the number of convolutional
kernels to arrive at the optimal production configuration. We
observed a consistently high accuracy of detection of infant
sleeping position changes from turning to facedown positions
with a potential towards even higher accuracies with caregiver
feedback for model retraining. Therefore, this system is a viable
candidate for consideration as a non-intrusive solution to assist
in preventing SIDS.
Index Terms—Sudden Infant Death Syndrome, SIDS, convo-
lutional neural network, image classification, posture estimation,
edge deep learning
I. I NTRODUCTION
Sudden Infant Death Syndrome (SIDS) is a leading cause of
death in infants. Research has led the scientific community to
believe that babies who succumb to SIDS are born with one or
more conditions that may lead to a fatality in response to what
are called internal and external stressors [1]. Researchers have
identified three primary risk factors that must all be present
and combine to result in the death of an infant from SIDS
[2]. The first of the risk factors is the infant being vulnerable.
Infants may be susceptible to an abnormality in the portion of
the brain that controls heart rate and respiration. The second
risk factor is the crucial nature of the developmental period of
the baby. For example, during the first six months of a baby’s
life, rapid growth may destabilize the baby’s internal systems.
The third and most important risk factor is the set of external
or outside stressors. These include things such as the sleeping
position of the baby, especially when the baby sleeps on the
stomach in a facedown position. All three risk factors listed
above must be present simultaneously for a SIDS fatality to
occur. It is recognized by experts that since the first two risk
978-1-6654-1252-0/21/$31.00 ©2021 IEEE
DOI 10.1109/SMARTCOMP52413.2021.00061
Authorized licensed use limited to: Koninklijke Philips N.V.. Downloaded on May 09,2022 at 08:35:52 UTC from IEEE Xplore. Restrictions apply.
factors stated above cannot be seen or pinpointed, the most
effective way to reduce the risk of SIDS is to reduce or remove
outside stressors altogether [1, 2]. This involves eliminating
the sleeping-position-related-risk by placing the sleeping baby
on the back instead of on the stomach and ensuring the baby
doesn’t turn from the back sleeping position to the stomach
sleeping position.
A number of tools have been created to attempt the task
of avoiding the sleeping position as an outside stressor. These
include baby sleeping sacks, breathing monitors, movement
sensor pads, special mattresses, bumpers, etc [3]. The U.S
Food and Drug Administration (FDA) has made it clear that
a safe sleep environment for a baby is for the infant to be
alone in the crib or bassinet, free of other objects including the
devices mentioned above [4]. Therefore, it would be beneficial
to create a non-intrusive, contact-less, and automated system
that can be used to ensure the baby maintains a back-sleeping
position while asleep. The system should have no devices,
objects, or sensors in contact with the baby in order to prevent
risks. The device should be expedient in detecting the sleeping
position outside stressor and should alert caregivers quickly so
that they can take immediate remedial action.
In this paper, we have proposed a custom, convolutional
neural network (CNN) based computer-vision system that can
process images from a camera focused on the baby’s bed
which can automatically detect the sleeping position of the
baby. Figure 1 presents the sleeping positions of the baby that
need to be detected. The system also detects an attempt by the
baby to change position from the back sleep position to the
stomach sleep position. The system automatically generates a
local audible alert and sends an alert notification to caregivers’
mobiles. The system allows the caregivers to provide a valida-
tion feedback regarding whether the alert is a valid alert or is a
false positive. This feedback can be used to retrain the system
so that accuracy levels are increased. The system is optimized
for low computational power and a low memory footprint,
which creates the potential for it to be embedded in edge
devices, such as certain baby monitors and home computers.
Our work highlights the potential to use deep learning
techniques at the edge to create non-intrusive life-saving solu-
tions that eliminate the need for manual monitoring. The deep
learning techniques can also eliminate the use of automated
monitoring devices, which, because they are intrusive, may
interfere with the objectives of such monitoring. We have
286
 Fig. 1. Sleeping position classifications.
demonstrated that deep learning techniques can be optimized
for settings that necessitate the use of low power and low
memory footprints. Deep learning techniques optimized for
low power and low memory footprint requirements would
open up the avenues for embedding machine learning in
edge devices, such as cameras, that are normally found in
clinical settings. Optimized deep learning techniques would
also expand the use of deep learning techniques into settings
where privacy concerns would hinder the transfer of images
from a clinical setting into the cloud for processing. Integration
of edge deep-learning models with mobile notification services
as demonstrated in our work creates the potential to free-up
the anxiety of caregivers while increasing the effectiveness of
care.
II. R ELATED W ORK
The use of Convolutional Neural Network (CNN) systems
for various computer vision tasks is well established, inspired
by the work of Hubel and Wiesel on the vision systems of
animals [5]. A progression of various advancements in the
architecture of CNNs can be found in [6-11]. Currently, one
of the primary uses of CNNs in computer vision is image
classification [12-15]. In image classification, the input images
are classified into one of a set of pre-specified categories. The
key aspect in using image classification CNNs for a certain
task is to formulate the task as a classification problem. A
related class of tasks is activity detection from videos [16-19]
wherein a specific activity, typically by a human, is detected
from a given video. An area of research related to baby
sleeping-position detection is that of human-pose estimation,
which is the localization of human joints in images or videos
and the search for a specific pose among all articulated poses.
Examples of prior research in human pose estimation include
[20-25].
We experimented with various prior classes of neural net-
work model architectures for the objectives of this work. Most
of the prior art deep-learning models for image classification
and posture estimation that we looked at required significant
computing resources as well as memory ranging from tens of
megabytes to hundreds of megabytes, precluding their use in
low power, low memory edge devices. Hosting these models
in the cloud creates cost and privacy challenges. These models
also take up to tens of seconds to process images. This would
render their use in time-critical tasks, such as quickly detecting
the SIDS outside stressor, impractical. Stripping-down prior art
models to make them edge-ready led to downgrades in their
performance. While techniques for patient monitoring using
sensors such as bed sensors and thermal sensors can be used
287
Authorized licensed use limited to: Koninklijke Philips N.V.. Downloaded on May 09,2022 at 08:35:52 UTC from IEEE Xplore. Restrictions apply.
Fig. 2. Custom CNN architecture.
to detect movement and temperature of the baby, they are not
useful in identifying the specific sleeping posture of the baby.
III. M ODEL
Since images of the head and shoulders area contain a
limited number of features, we hypothesized a simplified
custom convolutional neural network which implements image
classification for sleeping pose estimation. Then, we trained
the model on a substantial number of baby images in various
sleeping positions that would be suitable for accomplishing
the objectives of our work. The three classifications for the
image classification task were the following: (1) the baby
in a back sleeping position (normal); (2) the baby changing
from one position to another; (3) the baby sleeping in a
stomach sleeping position (outside stressor). Our proposed
architecture (see Figure 2) draws upon recent advancements
in image classification and pose estimation, but, at the same
time, simplifies these so as to address the challenges of
computational load and memory footprint at the edge.
In section III.A, we describe the components of our custom
CNN model. In section III.B, we describe the components that
surround the core custom CNN model in order to provide the
full functionality of a complete practical system. In section
III.C, we describe the process adopted to generate the training
and test datasets. The approach we took for evaluation of our
candidate models is presented in section III.D. Results of the
evaluation are presented in section III.E.
A. Model Architecture
The overall architecture of our custom CNN can be repre-
sented as follows:
Input image → [Convolutional Layer + Max Pooling Layer]
x multiple times → Fully Connected Layer(s) → Output.
1) Input Image: A 2D image of the baby captured by the
camera is the input data. The input image is converted into
grayscale and rescaled to 100 pixel x 100 pixel size. This
preprocessing to grayscale and resizing is performed both for
training data as well as for live images captured for sleeping
pose estimation. This preprocessing improves computational
efficiency and speed.
2) Convolutional Layer: For this system, a 3x3 convolu-
tional window was used. The output of the above convolution
is passed through an activation function. We used the Rectified
Linear Unit (ReLU) as the activation function due to its ability
to eliminate the issue of vanishing gradients in large scale

Fig. 3. End-to-end system design.
convolutional neural networks. Once the output is calculated,
the input window is shifted by the stride value. For this system,
a stride value of 1 was used. Multiple filters could be used
within a convolutional layer.
3) Max Pooling Layer: Max Pooling layer helps in picking
up the most relevant features and reducing the size of the data.
For this system, the window/kernel size of 2x2 with a stride
of 2 was used with Max (maximum of the given values) as the
pooling function. Max Pooling retains the largest value within
each window.
4) Dense Layer: Dense layers or fully connected layers
are the multilayer perceptron layers, where all inputs are
connected to the required outputs. In the final layer in our
implementation, the output is the set of probabilities of
three classes corresponding to the three possible sleeping
positions. In this particular application of dense layers, the
Sigmoid activation function was used. The final dense layer
had a dropout rate of 0.5 (50% probability of shutting down
randomly chosen nodes in each layer of the neural network
[26]).
The custom CNN system was built leveraging Python 3
with Jupyter Notebook as the development environment.
The core of the model was built using TensorFlow/Keras.
Sequential modeling (where layers are organized sequentially)
was used to develop the various layers in the neural network.
The open source code of our implementation can be found at:
https://github.com/viveksbharati/SIDS
B. Supporting components of end-to-end system
Figure 3 shows the high-level system functionality with
end-to-end system components including the above CNN
architecture at its core.
1) A camera is positioned facing the baby. The camera is
connected to a computer (which could be a processing
unit that is part of the camera). The program running on
the computer is capable of taking pictures.
2) The high resolution color image from the camera is
converted to grayscale, resized to 100 x 100 pixels, and
passed on to a Convolutional Neural Network whose
architecture was described earlier. The model returns
confidence levels against the three categories of sleeping
288
Authorized licensed use limited to: Koninklijke Philips N.V.. Downloaded on May 09,2022 at 08:35:52 UTC from IEEE Xplore. Restrictions apply.
positions (Sleeping on Back – OK , Baby trying to turn –
Alert, Baby turned over and lying on stomach – Alarm).
3) For a specific image captured, if the system decides the
baby is OK; it moves to Step 1 to capture the next image
and proceeds to further steps
4) In the case when the system decides to notify the
caregivers, an onscreen visible indicator and an audible
alert are switched on.
5) Also, the system can generate multiple channels of alerts
such as an SMS text message. These alerts are sent to
pre-configured mobile numbers. The alerts can contain
messages such as “The baby has turned over. Please
check on the baby.” In our implementation, we used
Amazon Simple Notification Service to send notifica-
tions.
6) Alternatively, a Push Notification can be sent to a mobile
application with the alert message along with the picture
from the camera. The system can generate multiple alerts
in multiple channels (for example, SMS to a few phones,
Push Notifications to multiple mobile devices, voice
messages on home automation devices, etc). The intent
is to ensure that a channel failure or latency does not
hinder the performance of the system.
7) The Push Notifications can be sent along with the
captured baby image with options for the caregiver to
acknowledge and validate the notification. Two options
are provided for the caregiver: a) “OK” to acknowledge
the notification as correct and switch off the alarm; b)
“False Alarm” to indicate that the system prediction is
incorrect.
8) Based on the caregiver’s acknowledgement that the
notification is correct, the corresponding images are an-
notated to the predicted class. If the caregiver disagrees
with the prediction, the prediction with the next higher
probability value is taken as the annotation for the next
round of training.
9) Periodic retraining with live images can also be per-
formed. As explained in Step 1, the system continuously
captures the baby’s images for predicting the baby’s
position. These images are retained for retraining along
with the predictions corresponding to these images. Any
image where there is an acknowledgement from the
caregiver (Step 8) is also included for the retraining
process. Periodic retraining of the model with the images
captured as part of the system execution adapts the
system to a specific baby.
C. Dataset generation
The datasets needed to train and test the model, which
would be a large number of images of sleeping babies in the
three positions discussed earlier, were not publicly available.
Therefore, we generated the datasets needed for this work.
Lifelike baby dolls of three different ethnicities in equal
proportions were used to generate the datasets. The continuous
stream of image data was provided by a webcam that was
trained on the head and shoulders area of the baby. OpenCV
was used to periodically capture images from the video feed to
create the datasets. While capturing the images, the angle of
the baby, position of the arms and the position of the baby
relative to the crib were changed continuously to account
for movements of the baby in the crib. Given the limited
set of features in the head and shoulders areas of the baby,
training of the model with datasets generated with realistic
baby dolls provides a starting point for the model, which
then gets enhanced with live images during periodic retraining
when deployed in live settings.
A total of 9,000 images generated as above were used as
seeds. Every seed image was also used to generate additional
images using image transformations. The ImageDataGener-
ator part of Keras image processing package was used for
rotating, cropping, and zooming the seed images. Additional
transformations involved altering the height and width of each
seed image and flipping the seed images horizontally. The
parameters for the transformations were:
• Image rotation with range up to 20 degrees
• Image shift (pan) horizontally/vertically by 10% of the
image width/height
• Horizontal flips
• Zoom range of 20%
• Up to 5% shearing of the image
This transformation process was carried out to ensure the
success of the neural network model in classifying different
sleeping positions regardless of any new positioning of the
camera, as well as variations in the image quality and camera
quality. Along with the seed and transformed images, the
total dataset included about 18,000 images across the three
sleeping position categories in nearly equal proportions. 5%
of that (about 900) was used as the test set, and the remaining
95% was used as the training set. Care was taken to avoid
including transformed images of training set images in the test
set. The images were then randomized and labeled to the three
categories (OK - Face Up, Alert - Turning, Alarm - Facedown).
All images were converted to grayscale, and scaled down to
100 x 100 pixels. The image size was chosen after trial and
error between the visible features of the baby’s position and
optimal size for model generation. The sized images were
pickled (utility to marshal and unmarshal a data structure)
for later use. The data was normalized using (x - x.min())
/ (x.max() - x.min()). In our case, since the grayscale image
had just one channel, we could normalize the grayscale image
data by dividing it by 255, changing every pixel value to be
between 0.0 and 1.0.
D. Evaluation
The combination of a Convolutional Layer with a Max
Pooling Layer was used as a single convolutional processing
unit. Different models were built and evaluated with five
different combinations of
• Number of (Convolutional layer + Max Pooling layer)
convolutional processing units
• Number of filters within each Convolutional layer
289
Authorized licensed use limited to: Koninklijke Philips N.V.. Downloaded on May 09,2022 at 08:35:52 UTC from IEEE Xplore. Restrictions apply.
• Number of Dense layers
The five models evaluated were:
• Model 1: 1 x (2D convolutional layer with 32 kernels +
1 Max Pooling layer) followed by 1 Dense layer.
• Model 2: 2 x (2D convolutional layer with 32 kernels +
1 Max Pooling layer) followed by 1 Dense layer.
• Model 3: 3 x (2D convolutional layer with 32 kernels +
1 Max Pooling layer) followed by 1 Dense layer.
• Model 4: 3 x (2D convolutional layer with 64 kernels +
1 Max Pooling layer) followed by 2 Dense layers.
• Model 5: 3 x (2D convolutional layer with 32 kernels +
1 Max Pooling layer) followed by 2 Dense layers.
The models were compiled with the following parameters:
• The loss function used was “Sparse Categorical Cross
Entropy”, which computes the cross entropy loss between
the labels and predictions.
• The Adam optimizer was used with learning rate = 0.001,
beta 1 = 0.9, beta 2 = 0.999, epsilon = 1e-07 and
amsgrad = False
• Capturing of the accuracy matrix was enabled using
TensorBoard
Accuracy is calculated by counting the correct predictions
and dividing it by the total number of predictions. The Sparse
Categorical Cross Entropy function was used to calculate the
loss. The model that yielded the best results, in terms of a
higher accuracy rate combined with lower error/loss as well
as lower memory and lower training time, was sought for
production deployment.
Accuracy evaluation: Model 1 never improved over an
accuracy of 0.4. The rest of the models yielded much higher
accuracy levels and tended to plateau around 10 epochs.
Therefore, the number of training epochs was reduced to
12. Figure 4 presents the results of the accuracy evaluation
for training. Figure 5 presents the results of the accuracy
evaluation for validation. Model 1 is not included in either
figure because of its significantly lower performance.
Fig. 4. Accuracy evaluation - Training.
 Fig. 5. Accuracy evaluation - Validation.
Loss evaluation: Model 1 had its lowest loss function of
0.5 after 14 epochs. For the other models, the loss function
considerably reduced after 10 epochs. Figure 6 presents the
results of the loss evaluation for training. Figure 7 presents
the results of the loss evaluation for validation. Since Model
1 had significantly poorer performance compared to the other
models, it is not included in either figure.
Fig. 6. Loss evaluation - Training.
290
Authorized licensed use limited to: Koninklijke Philips N.V.. Downloaded on May 09,2022 at 08:35:52 UTC from IEEE Xplore. Restrictions apply.
Fig. 7. Loss evaluation - Validation.
E. Results
Model 1 had a training time of 20 minutes. Since the
training and validation accuracies didn’t improve beyond 0.4
and loss didn’t reduce below 0.5, this configuration was
discarded.
Model 2 had good accuracy and loss performance. The size
of the model was 850 kB. The training time was 162 minutes.
Model 3 had the best accuracy and loss performance. The
size of the model was 500 kB. The training time was 25
minutes.
Model 4 had good accuracy and loss performance. The size
of the model was 5.7 MB. The training time was 52 minutes.
Model 5 had good accuracy and loss performance. The size
of the model was 1.5 MB. The training time was 25 minutes.
Model 3 had the best overall accuracy and loss performance,
along with the least memory footprint and training time.
Therefore, it was selected for production deployment.
During the live test runs, the system always predicted
the right posture, especially for the faceup and facedown
positions (i.e., the accuracy rates were close to 100% during
our observations). The processing time taken for the system
to capture three continuous images, transmit the data to the
model, classify all three images, perform majority voting on
the posture estimations, and generate corresponding notifica-
tions locally and on the mobile (in case of the baby turning
or lying facedown) was within 3 seconds.
IV. RUNTIME C ONFIGURATION O PTIONS
The production model could be deployed on low-powered
single-board computers, for example, Raspberry Pi or Nvidia
Jetson Nano. We have implemented the system on a laptop
with a connected webcam as well as on a Raspberry Pi device
with a miniature HD camera. The Raspberry Pi implementa-
tion setup was as follows:
 Hardware:
• Raspberry Pi 4 Model-B with 8 GB RAM, Quad-core
64-bit Broadcom 2711, Cortex A-72 processor
• OV5647 5 MP 1080p IR-cut camera mounted on the crib
and focused on baby’s head and shoulders area
Software:
• 64 bit Debian GNU/Linux OS installed on Pi 4. Camera
module enabled and Pi connected to WiFi.
• Python 3.7.3 for runtime
• Tensorflow 2.3.0/Keras 2.4.0 to load and run model
• OpenCV 4.1.1 for image capture and manipulations
• Boto3 - 1.17.30 to communicate with Amazon Web
Services Simple Notification Service
The model could be further fine-tuned and ported on to Ten-
sorFlow Lite to optimize the footprint and execution efficiency.
This would help in ensuring portability to a host of edge
platforms, including mobile phones (iOS/Android) as well
as platforms with embedded Linux and microcontroller-based
single-board computers based on ARM Cortex-M Series.
V. C ONCLUSION
We have proposed a custom CNN system to help prevent
the Sudden Infant Death Syndrome. It can provide alerts
via mobile notifications to help avoid the sleeping position
outside stressor. The system has been developed to minimize
computational load and memory footprint, which enables it to
be embedded in edge devices, such as baby monitors, low-end
home computers and low-cost single-board computers such
as Raspberry Pi devices. Our experiments show that a model
with three sequential sets of convolutional processing units,
each consisting of a 2D Convolutional layer with 32 kernels
and a Max Pooling layer, followed by one fully connected
Dense layer is able to estimate the sleeping pose of the
baby and any pose changes with high accuracy, and alert
caregivers on mobile devices in about three seconds. The
memory footprint of the model was around 500 kilobytes.
This gives us confidence that this custom convolutional neural
network model is a viable candidate for consideration as a non-
intrusive, contact-less, low-cost technology to help prevent
Sudden Infant Death Syndrome. Our work highlights the
potential to use deep learning techniques at the edge to create
non-intrusive life-saving solutions that eliminate the need for
manual monitoring.
R EFERENCES
[1] National Center of Child Health and Human Development, “Re-
search on Possible Causes of SIDS,” nichd.nih.gov, [Online]. Available:
https://bit.ly/2XHaKrb. [Accessed December 12, 2020].
[2] National Center of Child Health and Human Development, “What is
SIDS?,” nichd.nih.gov, [Online]. Available: https://bit.ly/3qgsaY5. [Ac-
cessed December 12, 2020].
[3] A. Wehrli, “SIDS: The 10 Best Prevention Products and 5 That
Were Recalled,” babygaga.com, May 26, 2018 [Online]. Available:
https://www.babygaga.com/sids-the-10-best-prevention-products-and-5-
that-were-recalled/. [Accessed December 12, 2020].
[4] U.S Food and Drug Administration, “Baby Products with
SIDS Prevention Claims,” fda.gov, October 17, 2019. [Online].
Available: https://www.fda.gov/medical-devices/products-and-medical-
procedures/baby-products-sids-prevention-claims. [Accessed December
12, 2020].
291
Authorized licensed use limited to: Koninklijke Philips N.V.. Downloaded on May 09,2022 at 08:35:52 UTC from IEEE Xplore. Restrictions apply.
[5] D. H. Hubel and T. N. Wiesel, “Receptive fields, binocular interaction,
and functional architecture in the cat’s visual cortex,” The Journal of
Physiology, vol. 160, pp. 106–154, 1962.
[6] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86, no.
11, pp. 2278–2323, 1998.
[7] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification
With Deep Convolutional Neural Networks,” In Proc. Advances in Neural
Information Processing 25, 2012, pp. 1097-1105.
[8] M. Zeiler and R. Fergus, “Visualizing and Understanding Convolutional
Networks,” In Proc. European Conference on Computer Vision 2014,
2014, pp. 818-833.
[9] C. Szegedy, W. Liu, Y. Jia, et al, “Going Deeper with Convolution,”
In Proc. IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), 2015, pp. 1-9.
[10] K. He, X. Zhang, S. Ren et al, “Deep Residual Learning for Image
Recognition”, In Proc. IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2016, pp. 770-778.
[11] F. Chollet, “Xception: Deeplearning with Depthwise Separable
Convolution,” arXiv.org, Apr. 2017. [Online]. Available:
https://arxiv.org/abs/1610.02357. [Accessed December 12, 2020].
[12] S. Soffer, A. Ben-Cohen, O. Shimon, et al., “Convolutional Neural
Networks for Radiologic Images,” Radiology, vol. 209, no. 3, pp. 590-
606. 2019.
[13] S. Ahlawat, A. Choudhary, A. Nayyar, et al., “Improved Handwriting
Recognition Using Convolutional Neural Networks,” Sensors 2020, vol.
20, no. 12, pp. 3344, 2020.
[14] M. Coskun, A. Ucar, O. Yildrium, et al., “Face recognition based
on convolutional neural network,” In Proc. International Conference on
Modern Electrical and Energy Systems (MEES), 2017, pp. 376-379.
[15] Z. Zhao and S, Xu, “Object Detection with Deep Learning: A Review,”
IEEE Transactions on Neural Networks and Learning Systems, vol. 30,
no. 11, pp. 3212-3232, 2019.
[16] A. S. Voulodimos, D. I. Kosmopoulos, N. D. Doulamis, and T. A.
Varvarigou, “A top-down event-driven approach for concurrent activity
recognition,” Multimedia Tools and Applications, vol. 69, no. 2, pp.
293–311, 2014.
[17] A. S. Voulodimos, N. D. Doulamis, D. I. Kosmopoulos, and T. A.
Varvarigou, “Improving multi-camera activity recognition by employing
neural network based readjustment,” Applied Artificial Intelligence, vol.
26, no. 1-2, pp. 97–118, 2012.
[18] T. Kautz, B. H. Groh, J. Hannink, U. Jensen, H. Strubberg, and B. M.
Eskofier, “Activity recognition in beach volleyball using a DEEp Con-
volutional Neural NETwork: leveraging the potential of DEEp Learning
in sports,” Data Mining and Knowledge Discovery, vol. 31, no. 6, pp.
1678–1705, 2017.
[19] K. Tang, B. Yao, L. Fei-Fei, and D. Koller, “Combining the right features
for complex event recognition,” In Proc. 2013 14th IEEE International
Conference on Computer Vision, 2013, pp. 2696–2703.
[20] A. Toshev and C.Szegedy, “DeepPose: Human pose estimation via
Deep Neural Networks,” arXiv.org, Aug. 2014. [Online]. Available:
https://arxiv.org/abs/1312.4659. [Accessed December 12, 2020].
[21] J. Tompson, R. Goroshin, A. Jain, et al., “Efficient object localization
using convolutional networks,” arXiv.org, June 2015. [Online]. Available:
https://arxiv.org/abs/1411.4280. [Accessed December 12, 2020]
[22] S. Wei, V. Ramakrishna, et al., “Convolutional Pose
Machines”, arXiv.org, Apr. 2016. [Online]. Available:
https://arxiv.org/abs/1602.00134. [Accessed December 12, 2020]
[23] J. Carreira, P. Agrawal, et al., “Human Pose Estimation with It-
erative Error Feedback”, arXiv.org, June 2016. [Online]. Available:
https://arxiv.org/abs/1507.06550. [Accessed December 12, 2020]
[24] A. Newell, K. Yang, J Deng, ”Stacked Hourglass Networks for
Human Pose Estimation”, arXiv.org, July 2016. [Online]. Available:
https://arxiv.org/abs/1603.06937. [Accessed December 12, 2020]
[25] Papandreou, George, et al. “PersonLab: Person Pose Estimation
and Instance Segmentation with a Bottom-Up, Part-Based, Geomet-
ric Embedding Model,” arXiv.org, Mar. 2018, [Online]. Available:
arxiv.org/abs/1803.08225. [Accessed December 12, 2020]
[26] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut-
dinov. 2014. “Dropout: A Simple Way To Prevent Neural Networks From
Overfitting,” Journal of Machine Learning Research, vol. 15, no. 56, pp.
1929–1958, 2014.