IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 3, September 2024, pp. 2666~2673

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i3.pp2666-2673  2666

Stand-off concealed firearm detection using motion tracking

and convolutional neural networks

Henry Muchiri Muriithi1, Ismail Lukandu Ateya1, Gregory Wanyembi2

1
School of Computing and Engineering Sciences, Strathmore University, Nairobi, Kenya
2
School of Computing and Informatics, Mount Kenya University, Nairobi, Kenya

Article Info ABSTRACT

Article history: The standoff detection of concealed firearms is crutial in managing public
security in public spaces. Currently employed standoff concealed weapon
Received Mar 23, 2023 detection techniques employ electromagnetic wave imaging which has been
Revised Feb 10, 2024 found to be extremely slow and may require expensive hardware and may
Accepted Feb 24, 2024 not be applicable in public open spaces. Inorder to maintain safety in open
spaces, artificial intelligence enabled video surveillance systems have been
widely adopted. This poses an opportunity to explore video surveillance
Keywords: cameras as concealed weapon detectors. A review of existing video
surveillance based automated weapon detection approaches discovered that
Automated concealed firearm
the focus was on the detection of unconcealed firearms leaving a gap in the
detection
detection of concealed firearms. This study addresses the aforementioned
Convolutional neural networks
gap by providing a standoff concealed firearm detection approach on video
Skeletal-based motion tracking
based on skeletal-based human motion tracking and convolutional neural
Smart surveillance
networks (CNN). The motion of armed and unarmed persons was tracked
using a depth camera and further classified using CNN model. The
developed model reported 100% accuracy, precision, and recall scores.
These results outperformed results obtained from traditional machine
learning models therefore highlighting the superior capability of the
proposed approach for concealed firearm detection on video to complement
the efforts of human video surveillance operators.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Henry Muchiri Muriithi
School of Computing and Engineering Sciences, Strathmore University
Ole Sangale Rd, off Langata Road, Madaraka Estate, Nairobi, Kenya
Email: [email protected]

1. INTRODUCTION
All over the world, populations are facing an increasing burden of firearm violence and mass
shootings on mortality and injuries [1]. Mortality from firearms contributes more than 500 deaths each day
worldwide with a majority of this deaths resulting from homicides [1]. An additional 2,000 people are injured
or maimed by gunshots every day [2]. With these alarming statistics, providing more significant control over
firearm usage is a crucial factor in reducing the effects of firearm violence. A challenging task for law
enforcement officers is the detection of concealed firearms [3]. Concealed weapon detection approaches can
broadly be categorized as either stop and search approaches or standoff approaches [3]. Stop and search
approach for example walk through metal detectors requires persons to stop at screening stations to be
searched while standoff approaches suspects are screened at a distance [3]. Stop and search approaches are
however only limited to entry points of buildings and therefore leaving out other areas such as open streets
where the public is also in danger of attacks from firearms. Standoff approaches on the converse can be
employed in open streets.

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  2667

Among the solutions employed for the standoff concealed weapon detection is the application of
electromagnetic wave imaging techniques or the use of video surveillance [3], [4]. The electromagnetic
techniques include the use of ultrasound, mmWave, Terahertz, infrared, fusion of visual RGB image and
infrared and X-ray imaging [5]. Various researchers have proposed electromagnetic based standoff concealed
weapon detection solutions [5]–[8]. Some researchers concur that the application of electromagnetic wave
imaging techniques have a long processing time and require expensive hardware that may not be possible in
many places such as open streets [4], [5].
Currently there are millions of video surveillance cameras installed in many open streets to maintain
security [9]. These systems rely on human operators to manna them and communicate to officers on the
ground incase suspicious activity or behavior is observed [9], [10]. Video surveillance systems however,
have the limitation that they require constant human supervision which is impractical especially in vast
volumes of data [4], [11]. An attractive alternative is the deployment of automated video surveillance
systems where potential criminal activites can be autonmously detected using artificial intelligence
techniques and prevent them before they occure [3], [10]. Various authors have proposed solutions for
intelligent standoff weapon detection techniques on video surveillance systems.
Ahmed et al. [9] implemented a real-time weapon detection approach using a scaled YOLOv4
object detector with the ability to detect unconcealed firearms with high mean average precison rates of over
92%. The approach achieved lower latency, higher throughput, and improved privacy by deploying on Jetson
Nano edge computing device. Narejo et al. [10] proposed a firearm detection technique using YOLOv3
object detection neural network with the ability to detect unconcealed firearms and subsequently sending an
alarm to security enforcement personnel. This approach outperformed intitial approaches that employed
YOLOv2 and traditional convolutional neural network (CNN) approaches. Bhatti et al. [11] developed a
firearm detection technique in real-time CCTV using YOLOV4 object detection neural network that could
detect unconcealed firearms in low resolution and brightness with over 91% average precision and F1 scores.
Figure 1 illustrates a sample detection outcome from the approach.

Figure 1. Demonstration of the detection of a unconcealed firearm by [11]

Sumi and Dey [12] developed an unconcealed firearm detector on video using YOLOv5 object
detection neural network. A comparative study revealed the models superior detection ability compared to
the baseline faster-RCNN based firearm detector approaches. Additionally, the study revealed that the
application of augumeted datasets yielded superior performance to non-augumented data.
The review of existing automated standoff firearm detection approaches on video surveillance
reveals great success and strides made in the detection of visible/unconcealed firearms. This however
presents a fundamental gap in the detection of concealed firearms on video surveillance. The detection of
concealed weapons people’s clothing is crutial in maintaining public safety [13].
This study aims to address the identified gap by presenting a real time automated standoff concealed
handgun detection by applying skeletal-based motion tracking technique to automatically track changes in
human motion on video as they walk with a concealed firearm tucked on their hip using state of the art deep
learning models. This approach is premised on the findings by [14], [15] which revealed that trained CCTV
human operators have the ability to identify concealed firearms because when a firearm is concealed into the
trouser pocket or the front waistband, it may hinder leg movements on that side of the body resulting in the
right stride being shorter than the left and a shorter arm swing. They attribute this disruption to the individual
attempts to either conceal the weapon or limit its movement so as not to drop it.
The proposed standoff surveillance solution would address the existing gap by enabling the
detection of concealed firearm on video surveillance while people are in motion for example on streets and
allow early detection before they are used to commit a crime. This approach would allow law enforcement

Stand-off concealed firearm detection using motion tracking and convolutional... (Henry Muchiri Muriithi)
2668  ISSN: 2252-8938

officers to rapidly and reliably screen individuals for concealed threats without any physical contact or
significant disruption of the suspect(s) activities and in effect reduce street crimes [5]. Automated motion
tracking has successfully been applied in various areas. In [16], [17] tracked motion for personality
assessment, [18], [19] tracked motion for gender classification and age estimation, [20], [21] tracked motion
for person identification and biometrics among others. To the best of our knowledge, there is no previous
work done to deploy motion tracking human pose estimation technique for concealed weapon detection on
video. The main contributions of this work are:
− Development of a novel 3D skeletal-based motion tracking dataset containing armed and unarmed
participants.
− Presentation of a concealed firearm detection approach on video using human motion tracking technique
and CNN techniques.
− Extend our previous study [22] that applied traditional machine learning algorithms for concealed firearm
detection.
The rest of the paper is organized as follows. Section 2 elaborates on the methods used to develop
the proposed solution. Section 3 contains the presentation of the research results and a discussion that
provides the reader with a deeper insight into the research findings. Finally, section 4 concludes the paper
and provides possible future directions in this area.

2. METHODS
In this section, detailed description and justification of the applied materials and methods set out to
achieve the study objectives.

2.1. Data collection instruments and procedure

The Microsoft Kinect depth camera together with the accompanying Kinect for windows software
development kit (SDK) was employed to track participants motion. Specifically, the Kinect v2 was used,
which enables capturing images at up to 30 frames per second (fps) and track 25 joints in the human body.
The camera captures the 3D spatial information of the tracked object relative to the camera. Microsoft Kinect
and other 3D sensing technologies offer real-time skeletal detection and tracking algorithms. Skeleton-based
motion tracking was preferred because they have lower data dimensionality in comparison with RGB-based
representations and therefore allowing the development of computationally faster and less complex systems
[23]. A Ceska handgun loaded with 11 rounds of ammunition was used. The pistol had a weight of 2.77
kilograms (kgs) and a length of 0.2 metres. A handgun was considered in this study due to its rampant use in
criminal activities [24].
The dataset creation process involved skeletal tracking and recording the 3D spatial-temporal
skeletal motion patterns of participants when armed and when unarmed. This was done in a lab environment
on a 6 metres long by 1-meter-wide clearly marked walkway using the Microsoft Kinect depth camera. The
sensor was elevated 1.2 metres above the ground and 1.5 metres away from one end of the walking path to
ensure that the subjects were within the range and field of view of the sensor for skeletal tracking. A physical
marker was placed near the end of the footpath, so that the subjects were aware of where they needed to stop
without having to look down to avoid skewing skew the results. Figure 2 illustrates the skeletal tracking of
one of the participants walking across the data collection pathway.

Figure 2. Data collection illustration

Int J Artif Intell, Vol. 13, No. 3, September 2024: 2666-2673

Int J Artif Intell ISSN: 2252-8938  2669

The recordings were made in two scenarios. In the first scenario, participants walked normally and
unarmed on the walkway at a self-selected speed. In the second scenario, the same participants were armed
with the Ceska handgun unholstered and concealed on the right hip. Each recording was about 3.2 seconds in
length and contained an average of 80 frames. All participants were informed of the study and signed the
required informed consent form. Data collection was approved by the Strathmore University Institutional
Ethics Review Committee (SU-IERC) and National Commission for Science, Technology and Innovation
(NACOSTI). To extract the spatial-temporal skeletal joint depth information from the recorded RGB-D
video, Kinect2 toolbox master application adopted from the works of [25] was used. The data contained the
tracked 3D skeletal joint position coordinates/point clouds.

2.2. Data encoding procedure

In the past few years, deep learning and in particular CNN have achieved state-of-the-art results in
image classification and object detection [11]. These neural networks however only work well on still images
and therefore exploiting them for spatial-temporal skeletal data analysis remains a challenge that requires the
encoding of data into an applicable form. To overcome this challenge, the study adopted the image encoding
approach proposed by [26]. The approach encodes the obtained spatial- temporal skeletal sequences into 2D
RGB image structure.
The encoding process of a skeleton sequence with N frames takes place by normalizing each 3D
joint coordinate (x, y, and z) in a given frame (f) into the range of 0–255, by applying in (1):

𝑘(𝑓)′ = 255 ∗ 𝑘(𝑓) − 𝑚𝑖𝑛{ 𝑐}/ 𝑚𝑎𝑥{ 𝑐} − 𝑚𝑖𝑛{ 𝑐} (1)

Where k is the coordinate (x, y, or z) of the tracked and recorded joint data, and min{c} and max{c} are the
minimum and maximum values of all coordinates in the sequence, respectively. The resultant encoded image
contains the (R, G, B) of a colour pixel which have been transformed from skeleton joints coordinates
(x, y, z)—x = R; y = G; z = B. A sample encoded RGB image representing the motion changes over time of
one tracked skeletal joint in this case joint number 4 is presented in Figure 3. The encoded images formed the
input to the CNN. The ratio of training set over validation set was set at 80%-20% resulting in 510 training
frames instances and 90 frames for validation.

Figure 3. Encoded RGB image

2.3. Convolutional neural network model architecture

The deployed CNN classifier was implemented using Python programming language, Keras
framework, and TensorFlow as the back-end. To achieve the much-required computational efficiency of the
model, all the image representations were reshaped to 9×9 pixels. Adam optimizer was used with the default
parameters.
The neural network architecture consisted of three convolutional layers, each with a 3×3 kernel, as
well as padding p=1 and stride S=1. This was followed by batch normalization layer and the rectified linear
unit (ReLU) as the activation function in each convolutional layer. To prevent model overfitting, an L2
kernel regularizer (L2(0.1) and a dropout layer (dropout (0.5) were implemented in each layer. This was
followed by a flattening layer to convert the output from the convolutional layers into a single feature vector
for classification. Lastly, two dense layers were placed, Dense (512) with ReLU activation function for
feature selection and dense (4) with sigmoid activation to achieve the binary image classification task.

3. RESULTS AND DISCUSSIONS

This study set out to improve the existing concealed firearm detection approaches by analyzing
encoded skeletal-based motion data using CNN. This section enumerates the classification results of the
developed model and additionally provides a detailed interpretation and discussion of the results. As a
classification machine learning problem, the CNN was expected to accurately distinguish between motion
images representations of armed and unarmed participants.

Stand-off concealed firearm detection using motion tracking and convolutional... (Henry Muchiri Muriithi)
2670  ISSN: 2252-8938

To measure the classification performance of the developed model, the study employed various
complementary metrics that were captured during model training and during testing and validation phases
[27]. The models learning accuracy, and loss function across the training and testing cycle was plotted
inorder to diagnose the models behavious. During testing and validation phase the confusion matrix and
presition, recall and F1 scores were measured.
Figure 4 provides a comparativee plot of the training and validation loss function of the model
across the 15 epochs. This plot is benefitial in indicating the overall behavior of the developed neural
network. The plot depicts a good fit characterized by a training and validation loss which both decrease to a
point of stability with a small gap between the two metrics [27]. This result is an indication that the model is
not overfitting or underfitting the data and therefore impying that the developed model was adequately
learninig and was able to generalize well.

Figure 4. Training and validation loss plot

The developed model reported a training and validation accuracies of 100% over 15 epochs as
illustrated by Figure 5. A significantly small gap between the training accuracy and validation accuracy
signifies a good model fit which is devoid of under or overfitting. A validation accuracy of 100% which is
obtained without overfitting denotes the superior capability of the developed model to accurately classify
motion images of armed and unarmed.

Figure 5. Training and validation accuracy plot

To further evaluate and understand the performance of the developed concealed fiream detection
CNN, the study employed the confusion matrix performance measurement tool. The obtained results are
presented in Figure 6. The presented results further affirm the superior performance of the model with zero
classification errors reported as false positives (FP) or false negatives (FN). This is an indication that the
developed model can accurately classify armed data (true positives (TP)) and distinguish that from images of
unarmed instances (true negatives (TN)).

Int J Artif Intell, Vol. 13, No. 3, September 2024: 2666-2673

Int J Artif Intell ISSN: 2252-8938  2671

Figure 6. Armed or unarmed confusion matrix

Because the developed model was a classification model, precision and recall performance metrics
were additionally employed [27]. The scores were compared with those of our previous study [22] that
employed traditional machine learning algorithms. The deep learning model presented 100% scores in all the
metrix in comparison to traditional machine learning algorithms which reported a maximum of 93% in all
measured matrix. The comparative results are presented in Table 1.

Table 1. Comparative analysis of the results

Algorithms Evaluation metrics
Accuracy (%) Precision (%) Recall (%)
Developed CNN 100 100 100
k-nearest neighbour 93.6 93.4 93.6
Random forest 90.6 90.7 90.6
Decision tree 68.60 68.60 68.60
Support vector machine 41.1 40.9 41.1
naïve Bayes 37.5 39.2 37.5

These results further confirm the superiority of the deep learning approach. The presented precision
scores indicate the correctness of the model to classify an image as armed with no cases of FP while the
presented recall scores indicate that the neural network’s ability to correctly classify all positive/armed cases
which in this case are the armed images. Additionally, the study was keen on measuring the detection time.
The video data used to train this model contained 81 frames recorded in about 4 seconds for a distance of
about 6 meters. The time of 4 seconds can be interpreted as the average walking time required for the
concealed firearm detection model to analyze motion and detect the fiream. This time is short and acceptable
for the detection task.

4. CONCLUSION
This study aimed to develop an efficient deep learning model for the standoff detection of people
carrying concealed firearms by tracking their motion. The model’s learning curve was evaluated together
with classification performance metrics such as accuracy, precision, and recall scores. The analysis of the
learning curve shows a good fit learning curve which indicates the model is able to satisfacrorily learn the
distinguishing features in the image inputs. The model presented an accuracy, precision, and recall scores of
100%. This finding is a confirmation of the ability of the developed CNN model to accurately classify armed
and unarmed images by 3D motion analysis. Therefore, this research concludes that it is possible to a great
extent to apply automated motion analysis using state-of-the-art video analysis techniques for concealed
firearm detection. This approach to automated concealed firearm detection by motion analysis is novel and a
great breakthrough in this area and will go along way in aiding the fight against crimes involving firearms.
The data applied in the study consisted of single persons walking towards the detection sensor/camera.
Following the outstanding performance presented by the developed model, future studies can apply data
consisting of multiple persons walking in different directions from the detection sensor/camera. This
extension to the study will be a great effort in mimicking real-life surveillance environments. Additionally,

Stand-off concealed firearm detection using motion tracking and convolutional... (Henry Muchiri Muriithi)
2672  ISSN: 2252-8938

the study was focused in an indoor lab environment, and as such future studies can focus on application of
the approach in real-time cctv footage.

ACKNOWLEDGEMENTS
We thank the National Research Fund (NRF) Kenya for the research grant- [Postgraduate grant
2016/2017], Kenya Police Service for providing the firearm that was used in this study, the 27 study
participants for their assistance in data acquising and research assistants for their support in cleaning and
pre-processing the data.

REFERENCES
[1] M. Werbick, I. Bari, N. Paichadze, and A. A. Hyder, “Firearm violence: a neglected ‘global health’ issue,” Globalization and
Health, vol. 17, no. 1, 2021, doi: 10.1186/s12992-021-00771-8.
[2] I. Cardoza, J. P. G. -Vázquez, A. D. -Ramírez, and V. Q. -Rosas, “Convolutional neural networks hyperparameter tunning for
classifying firearms on images,” Applied Artificial Intelligence, vol. 36, no. 1, Apr. 2022, doi: 10.1080/08839514.2022.2058165.
[3] A. Agurto, Y. Li, G. Y. Tian, N. Bowring, and S. Lockwood, “A review of concealed weapon detection and research in
perspective,” in 2007 IEEE International Conference on Networking, Sensing and Control, ICNSC’07, IEEE, 2007, pp. 443–448,
doi: 10.1109/ICNSC.2007.372819.
[4] S. A. A. Shah, M. A. Al-Khasawneh, and M. I. Uddin, “Review of weapon detection techniques within the scope of street-
crimes,” in 2021 2nd International Conference on Smart Computing and Electronic Enterprise: Ubiquitous, Adaptive, and
Sustainable Computing Solutions for New Normal, ICSCEE 2021, 2021, pp. 26–37, doi: 10.1109/ICSCEE50312.2021.9498007.
[5] X. Gao, H. Liu, S. Roy, G. Xing, A. Alansari, and Y. Luo, “Learning to detect open carry and concealed object with 77 GHz
radar,” IEEE Journal on Selected Topics in Signal Processing, vol. 16, no. 4, pp. 791–803, 2022, doi:
10.1109/JSTSP.2022.3171168.
[6] Z. Zhang, X. Di, Y. Xu, and J. Tian, “Concealed dangerous object detection based on a 77GHz radar,” in 2018 IEEE International
Workshop on Electromagnetics: Applications and Student Innovation Competition, iWEM 2018, IEEE, 2018, doi:
10.1109/iWEM.2018.8536660.
[7] Y. Li, Z. Peng, R. Pal, and C. Li, “Potential active shooter detection based on radar micro-doppler and range-doppler analysis
using artificial neural network,” IEEE Sensors Journal, vol. 19, no. 3, pp. 1052–1063, 2019, doi: 10.1109/JSEN.2018.2879223.
[8] C. V. Nelson, “Wide-area metal detection system for crowd screening,” in Sensors, and Command, Control, Communications,
and Intelligence (C3I) Technologies for Homeland Defense and Law Enforcement, vol. 5071, Sep. 2003, pp. 380-387, doi:
10.1117/12.484846.
[9] S. Ahmed, M. T. Bhatti, M. G. Khan, B. Lövström, and M. Shahid, “Development and optimization of deep learning models for
weapon detection in surveillance videos,” Applied Sciences, vol. 12, no. 12, 2022, doi: 10.3390/app12125772.
[10] S. Narejo, B. Pandey, D. E. Vargas, C. Rodriguez, and M. R. Anjum, “Weapon detection using YOLO V3 for smart surveillance
system,” Mathematical Problems in Engineering, vol. 2021, pp. 1–9, 2021, doi: 10.1155/2021/9975700.
[11] M. T. Bhatti, M. G. Khan, M. Aslam, and M. J. Fiaz, “Weapon detection in real-time CCTV videos using deep learning,” IEEE
Access, vol. 9, pp. 34366–34382, 2021, doi: 10.1109/ACCESS.2021.3059170.
[12] L. Sumi and S. Dey, “YOLOv5-based weapon detection systems with data augmentation,” International Journal of Computers
and Applications, vol. 45, no. 4, pp. 288–296, 2023, doi: 10.1080/1206212X.2023.2182966.
[13] L. Pang, H. Liu, Y. Chen, and J. Miao, “Real-time concealed object detection from passive millimeter wave images based on the
YOLOV3 algorithm,” Sensors, vol. 20, no. 6, 2020, doi: 10.3390/s20061678.
[14] I. T. Darker, A. G. Gale, and A. Blechko, “CCTV as an automated sensor for firearms detection: human-derived performance as a
precursor to automatic recognition,” in Unmanned/Unattended Sensors and Sensor Networks V, 2008, doi: 10.1117/12.800264.
[15] N. Meehan, C. Strange, and A. Garinther, “It’s the walk, not the talk: behavioral indicators of concealed and unholstered firearms
carrying,” Police Journal, vol. 94, no. 4, pp. 462–480, Sep. 2021, doi: 10.1177/0032258X20960777.
[16] J. Sun et al., “Relationship between personality and gait: Predicting personality with gait features,” in Proceedings - 2018 IEEE
International Conference on Bioinformatics and Biomedicine, BIBM 2018, IEEE, 2019, pp. 1227–1231, doi:
10.1109/BIBM.2018.8621300.
[17] L. Satchell, P. Morris, C. Mills, L. O’Reilly, P. Marshman, and L. Akehurst, “Evidence of big five and aggressive personalities in
gait biomechanics,” Journal of Nonverbal Behavior, vol. 41, no. 1, pp. 35–44, Sep. 2017, doi: 10.1007/s10919-016-0240-1.
[18] C. Xu et al., “Real-time gait-based age estimation and gender classification from a single image,” in Proceedings - 2021 IEEE
Winter Conference on Applications of Computer Vision, WACV 2021, IEEE, 2021, pp. 3459–3469, doi:
10.1109/WACV48630.2021.00350.
[19] L. Igual, À. Lapedriza, and R. Borràs, “Robust gait-based gender classification using depth cameras,” Eurasip Journal on Image
and Video Processing, vol. 2013, no. 1, 2013, doi: 10.1186/1687-5281-2013-1.
[20] H. Yamada, J. Ahn, O. M. Mozos, Y. Iwashita, and R. Kurazume, “Gait-based person identification using 3D LiDAR and long
short-term memory deep networks,” Advanced Robotics, vol. 34, no. 18, pp. 1201–1211, 2020, doi:
10.1080/01691864.2020.1793812.
[21] C. Á. -Aparicio, Á. M. G. -Higueras, M. Á. G. -Santamarta, A. C. -Vega, V. Matellán, and C. F. -Llamas, “Biometric recognition
through gait analysis,” Scientific Reports, vol. 12, no. 1, 2022, doi: 10.1038/s41598-022-18806-4.
[22] H. M. Muriithi, I. A. Lukandu, and G. W. Wanyembi, “Determining the location of a concealed handgun on the human body
using marker-less gait analysis and machine learning,” in 2nd International Conference on Next Generation Computing
Applications 2019, NextComp 2019, IEEE, pp. 1-6, Sep. 2019, doi: 10.1109/NEXTCOMP.2019.8883635.
[23] V. Silva, F. Soares, C. P. Leão, J. S. Esteves, and G. Vercelli, “Skeleton driven action recognition using an image-based spatial-
temporal representation and convolution neural network,” Sensors, vol. 21, no. 13, 2021, doi: 10.3390/s21134342.
[24] K. Nettles, C. Ford, and P. A. Prada-Tiedemann, “Development of profiling methods for contraband firearm volatile odor
signatures,” Frontiers in Analytical Science, vol. 1, 2022, doi: 10.3389/frans.2021.785271.
[25] Y. Zhu, Y. Zhao, and S. C. Zhu, “Understanding tools: task-oriented object modeling, learning and recognition,” in Proceedings
of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, IEEE, 2015, pp. 2855–2864, doi:

Int J Artif Intell, Vol. 13, No. 3, September 2024: 2666-2673

Int J Artif Intell ISSN: 2252-8938  2673

10.1109/CVPR.2015.7298903.
[26] H. H. Pham, L. Khoudour, A. Crouzil, P. Zegers, and S. A. Velastin, “Exploiting deep residual networks for human action
recognition from skeletal data,” Computer Vision and Image Understanding, vol. 170, pp. 51–66, 2018, doi:
10.1016/j.cviu.2018.03.003.
[27] M. Z. Naser and A. H. Alavi, “Error metrics and performance fitness indicators for artificial intelligence and machine learni ng in
engineering and sciences,” Architecture, Structures and Construction, vol. 3, no. 4, pp. 499–517, Nov. 2023, doi:
10.1007/s44150-021-00015-8.

BIOGRAPHIES OF AUTHORS

Dr. Henry Muchiri Muriithi holds a Ph.D. in Information Systems from

Strathmore university. He is Currently a lecturer and the Academic Director at the
Strathmore University School of Computing and Engineering Sciences (SCES). He
has been involved in several research projects notably the Microsoft Wise4Africa
Project 2019 and the Clinton Health Access Initiative (CHAI) – Drug Access Project
2015. He is also a reviewer for conferences and journals. He has presented conference
papers and published articles in various national and international reputable journals
of information systems and artificial intelligence. He can be contacted at email:
[email protected].

Prof. Ismail Lukandu Ateya Professor Ismail Ateya earned his Doctor
of Science degree in Applied Geophysics from Kyoto University, Japan in 2003 and
Graduate Diploma in Computer Science from University of Auckland, New Zealand
in 2006.He is currently the Director, Office of Faculty Affairs at Strathmore
University and lecturer at Strathmore University School of Computing and
Engineering Science. His research interests are on software and database modelling of
large information systems – and he has been an active member of the Database
Research Group. He can be contacted at email: [email protected].

Prof. Gregory Wanyembi holds a doctorate in management of ICTs

from Delft University of Technology, The Netherlands, in 2002. He has published
widely in the area of management and strategic utilization of ICTs in developing
countries. He is engaged in several research projects and supervision of postgraduate
students at both masters and Ph.D. from different universities in Kenya. He is a
reviewer of journal articles. He is professor at Mount Kenya Universiry, Kenya. He
can be contacted at email: [email protected].

Stand-off concealed firearm detection using motion tracking and convolutional... (Henry Muchiri Muriithi)