Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

Enhancing the Intelligence of Real-Time

Video Surveillance Systems with Automated
Anomaly Detection and Response
(Area of Focus: Security Organs)

NDIZEYE Maurice1; Dr. Emmanuel BUGINGO2

University of Kigali

Publication Date: 2025/03/04

Abstract: This study addresses the critical need for advanced threat detection systems by enhancing real-time video
surveillance with automated anomaly detection and response. It integrates machine learning algorithms for object and
activity recognition, leveraging AI and parallel computing for efficient and accurate anomaly detection. The research
evaluates the adaptability and effectiveness of machine learning models in diverse environments, emphasizing robust
response protocols for identified anomalies.

Key findings highlight challenges in identifying suspects amidst distractions like vehicles, animals, and crowds.
Advanced preprocessing and machine learning techniques significantly improved detection accuracy from 70.4% to 93%.
Parallel computing reduced model training time from 71 minutes to just 8 minutes, showcasing its efficiency. The study
recommends cost-effective solutions like building high-performance computing (HPC) clusters using Raspberry Pi nodes
for training and employing Hadoop clusters for video stream processing. Deploying optimized models on scalable web
servers enables real-time diagnostics and anomaly alerts, empowering security operators with faster and more actionable
insights.

These contributions advance the development of intelligent, efficient, and scalable surveillance systems that address
modern security challenges. The proposed methods improve decision-making and real-time responsiveness, providing a
significant leap forward in video surveillance technology.

How to Cite: NDIZEYE Maurice; Dr. Emmanuel BUGINGO. (2025). Enhancing the Intelligence of Real-Time Video Surveillance
Systems with Automated Anomaly Detection and Response (Area of Focus: Security Organs). International Journal of Innovative
Science and Research Technology, 10(2), 1217-1225. https://doi.org/10.5281/zenodo.14959415.

I. INTRODUCTION II. METHODOLOGY

The rise of global security concerns has highlighted the A. Data Collection and Preprocessing
limitations of traditional surveillance systems, which rely
heavily on manual monitoring and delayed responses. These  Data Sources: Video datasets were collected from online
systems are often reactive, with their primary function being websites and YAKIN Technology Ltd.'s surveillance
the provision of evidence after an incident rather than system, including recordings from public spaces, offices,
proactive threat detection and response. and parking areas.
 Preprocessing: Video streams were preprocessed to
Recent advancements in artificial intelligence (AI) and improve model performance by:
computer vision offer transformative potential in the security
domain. This paper explores the development of an intelligent  Resizing frames to standardized dimensions for
real-time video surveillance system capable of autonomously uniformity.
identifying and responding to potential threats, including  Extracting keyframes to reduce redundant data.
theft, intrusion, vandalism, and violence. The goal is to  Enhancing image quality through noise reduction,
leverage deep learning models and computer vision histogram equalization, and contrast adjustment.
techniques to transform passive surveillance into an active
and responsive security solution.

IJISRT25FEB564 www.ijisrt.com 1217

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

B. Feature Extraction and Selection F. Performance Evaluation

 Spatial features (object shapes and sizes) were extracted  The System was Tested on the Surveillance Infrastructure
using Convolutional Neural Networks (CNNs). of YAKIN Technology Ltd. Metrics Included:
 Temporal patterns (movement sequences) were analyzed
using Recurrent Neural Networks (RNNs) with Long  Accuracy: Improved anomaly detection accuracy from
Short-Term Memory (LSTM) units. 70.4% to 93%.
 Principal Component Analysis (PCA) was applied to  Response Time: Reduced detection and alerting times to
reduce dimensionality, retaining only the most relevant under 2 seconds.
features.  Scalability: Successfully processed multiple video
streams in real-time without system lag
C. Parallel Computing for Training Efficiency  Comparative analyses were conducted against existing
systems to validate improvements.
 Parallel computing techniques were applied using high-  User feedback was incorporated to fine-tune the system,
performance computing (HPC) clusters. These included: ensuring continuous improvement.
 Building affordable HPC clusters with Raspberry Pi
nodes. G. Research Design
 Utilizing GPU acceleration reduces model training time The research design for enhancing the intelligence of a
from 71 minutes to 8 minutes. real-time video surveillance system integrates multiple
components to ensure efficient processing, anomaly
D. Model Development detection, and alert generation. The system begins with a
video source (e.g., surveillance cameras) that streams live
 Deep Learning Models: The system utilized CNNs for footage into a Hadoop cluster for processing and storage.
spatial anomaly detection (e.g., unauthorized access) and The Hadoop cluster handles large-scale video data,
RNNs for temporal anomaly recognition (e.g., suspicious preprocessing it by extracting keyframes, resizing, and
movements). enhancing quality for further analysis. Preprocessed data is
 Integration of Adaptive Learning: A self-supervised then transferred to a High-Performance Computing (HPC)
learning mechanism allowed the system to adapt to cluster, which utilizes parallel computing and GPU
evolving behaviors and environmental conditions. acceleration to efficiently train deep learning models. These
 Hyperparameter Optimization: Techniques such as models, built using AI with Python, employ CNNs for
grid search and random search were employed to tune. spatial anomaly detection and RNNs for analyzing temporal
patterns and identifying suspicious movements or behaviors
E. System Deployment and Real-Time Processing in real time. When an anomaly is detected, the AI system
generates alerts, triggering alarms or sending notifications to
 Infrastructure: Optimized models were deployed on security operators for immediate action. The system operates
scalable web servers capable of processing live video continuously, leveraging adaptive learning to refine its
streams. accuracy over time (Ali, 2022). This design ensures real-time
 Video Stream Processing: Hadoop clusters were used to responsiveness, scalable data handling, and precise anomaly
handle large volumes of video data, enabling real-time detection, making it ideal for diverse surveillance
processing and anomaly detection across multiple environments. The flow chart visualizes the seamless data
cameras simultaneously. flow between components, emphasizing their interconnection
and role in achieving the system's objectives (Seemantula
 Model Quantization: To optimize resource usage,
Nischal, 2023).
models were quantized to reduce computational
requirements while maintaining high detection accuracy.

IJISRT25FEB564 www.ijisrt.com 1218

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

Fig 1: Flowchart of the System

IJISRT25FEB564 www.ijisrt.com 1219

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

III. CONCEPTUAL FRAMEWORK

Fig 2: Conceptual Framework

IV. DATA PRESENTATION, ANALYSIS AND situation, parallel computing affects the effectiveness and
INTERPRETATION OF FINDINGS precision of deep learning models.

The results of the study on the impact of parallel A. Distribution of Classes

processing and computing on developing deep learning To begin, we analyzed the distribution of different
models for anomaly detection and identification are presented classes (person, suspect, vehicle, and animal) in the dataset.
in this chapter. It contains a thorough examination of the This helps in understanding the balance of the dataset and the
information and findings from numerous experiments. The representation of the class.
goal is to give a thorough illustration of how, in this particular

IJISRT25FEB564 www.ijisrt.com 1220

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

Fig 3: Distribution of Images across Classes

B. Parallel vs Serial Computing sequentially by a single processor. The figure below visually
In the context of training deep learning models for compares training time reduction when parallel computing is
anomaly detection, we compared the performance of parallel applied to various tasks, demonstrating its efficiency in
and serial computing approaches. Parallel computing, which managing larger datasets and more complex models. This
uses multiple processors to perform computations comparison highlights the key benefits of parallel processing,
concurrently, significantly reduces model training times including reduced latency and faster convergence during
compared to serial computing, where tasks are processed training.

Fig 4: Parallel Computing vs Serial Computing

IJISRT25FEB564 www.ijisrt.com 1221

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

We used both serial and parallel computing techniques We found that the deep learning model was taking about
to train the identical deep learning model to assess how eight minutes to train when using parallel computing. By
parallel computing affected training time. The time used for dividing up the calculations among several processors, this
parallel computing (High-Performance computing) for model efficiency was attained, enabling the processing of tasks and
training is 0.126 hours, which is equal to 7.56 minutes, as data at the same time.
shown by the figure.
The following is a comparison of the training times that
The time used for Serial computing or model training is were noted.
equal to 1.174 hours, as shown by the figure, which is equal
to 70.44 minutes.

 *With Parallel Computing Training Time: 0.126 (h)  8

mins
 *With Serial Computing Training Time: 1.347 (h)  81
mins

Fig 5: Model Training Time Comparison

C. Model Performance Testing Findings Five times out of the six tests, the model accurately predicted
The great performance of the model's predictions the anomaly detection. This shows that, in the majority of
against the actual labels of the person samples during the circumstances, the model is highly accurate at finding the
testing phase for our model's performance on unseen photos. right abnormality.

IJISRT25FEB564 www.ijisrt.com 1222

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

Fig 6: Model Performance Testing Findings

The below figure shows the web interface used for a and alarm generated as a response to the operator or decision
real-time video surveillance system with the integration on maker.
the Artificial Intelligence with automated anomaly detection

Fig 7: Anomaly Detected in Real-Time with Alert

D. Related Work Look Once) for real-time object detection and RNN-based
Research in automated video surveillance has gained systems for behavior analysis. However, these approaches
significant traction in recent years. Studies have explored often face challenges such as computational inefficiency,
applications of convolutional neural networks (CNNs) for false-positive detections, and difficulty in generalizing across
object detection, human action recognition, and anomaly diverse environments. This research aims to build upon these
detection. Notable frameworks include YOLO (You Only advancements and overcome existing limitations by

IJISRT25FEB564 www.ijisrt.com 1223

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

integrating cutting-edge algorithms with practical [7]. Brown, S. (2021, Apr 21). MIT. Retrieved from
implementation strategies, including parallel computing and mitsloan.mit.edu: https://mitsloan.mit.edu/ideas-
model quantization. made-to-matter/machine-learning-explained
[8]. Brownlee, J. (2019, September 16). A Gentle
E. Limitations and Challenges Introduction to Transfer Learning for Deep Learning.
The project faced several limitations and challenges, Retrieved from https://machinelearningmastery.com/:
including the complexity of real-time video stream https://machinelearningmastery.com/transfer-
processing with low latency, ensuring accurate anomaly learning-for-deep-learning/
detection through machine learning, and integrating parallel [9]. Brownlee, J. (2023). Logistic Regression for Machine
computing with a Hadoop cluster for efficient data processing Learning. Machine Learning Mastery, 1-7.
and storage. Scaling the system to handle varied [10]. Brownlee, J. (2023, November 16). Machine Learning
environments and minimizing false alarms to avoid operator Mastery. Retrieved from
overload were also significant hurdles. Additionally, machinelearningmastery.com:
managing the infrastructure requirements for large-scale data https://machinelearningmastery.com/what-is-deep-
storage and processing, as well as addressing data privacy and learning/
security concerns, posed further challenges during both [11]. Dalibor Klusacek, G. P. (2023, May 19). Job
implementation and design. Scheduling Strategies for Parallel Processing.
Retrieved from Job Scheduling Strategies for Parallel
V. CONCLUSION Processing:
https://www.google.rw/books/edition/Job_Schedulin
This research integrated automated anomaly detection g_Strategies_for_Parallel_P/M1DXEAAAQBAJ?hl=
and reaction mechanisms into real-time video surveillance en&gbpv=1
systems using advanced machine learning models. It [12]. Dobbin, J. (2019, February 24). GPU vs CPU: What
significantly improved system accuracy and versatility, Matters Most for PC Gaming? Retrieved from
enabling efficient anomaly detection and rapid response. The https://www.hp.com/: https://www.hp.com/us-
system proved robust across different environments, en/shop/tech-takes/gpu-vs-cpu-for-pc-gaming
demonstrating high accuracy in real-time video feeds. [13]. Eremenko, K. (2018). Deep Learning A-Z™:
Parallel computing accelerated the development of deep Convolutional Neural Networks (CNN) - Step 3:
learning models, increasing classification accuracy to 0.93 Flattening. Flattening, 3-5.
from 0.904 and reducing training time from 71 to 8 minutes. [14]. Ethier, S. (2022, October 19). Parallel Computing:
This enhanced efficiency allowed for faster model iterations Intro to MPI. Retrieved from
and more robust results. The solution, deployable on a web https://researchcomputing.princeton.edu/:
application, empowers security operators with real-time https://researchcomputing.princeton.edu/sites/g/files/t
surveillance insights, improving security monitoring and oruqf311/files/documents/MPI_tutorial_Fall_Break_
response times. 2022.pdf
[15]. Ethier, S. (2022, October 19). Parallel Computing:
REFERENCES Intro to MPI. Retrieved from
https://researchcomputing.princeton.edu/:
[1]. Ahmed Elmetwally, R. E. (2024). Deep learning based https://researchcomputing.princeton.edu/sites/g/files/t
anomaly detection in real-time video. Deep learning oruqf311/files/documents/MPI_tutorial_Fall_Break_
based anomaly detection in real-time video, 1705- 2022.pdf
1716. [16]. Gandhi, R. (2018, May 27). Introduction to Machine
[2]. Ali, M. M. (2022). Real-time video anomaly detection Learning Algorithms: Linear Regression. Retrieved
for smart surveillance. Real-time video anomaly from towardsdatascience.com:
detection for smart surveillance, 1375-1388. https://towardsdatascience.com/introduction-to-
[3]. Ali, M. M. (2022). Real-time video anomaly detection machine-learning-algorithms-linear-regression-
for smart surveillance. Real-time video anomaly 14c4e325882a
detection for smart surveillance, 1375-1388. [17]. geeksforgeeks. (2024, Sep 04).
[4]. Ankush Balaram Pawar1Gayatri Shitole2, S. N. https://www.geeksforgeeks.org. Retrieved from
(2024). Intelligence Video Surveillance Using Deep https://www.geeksforgeeks.org:
Learning . Software Computing and Testing, 2456- https://www.geeksforgeeks.org/cnn-introduction-to-
2351. pooling-layer/
[5]. Bansod, S. (2019). Transfer Learning for video [18]. geeksforgeeks.org. (2024, Sep 04).
anomaly detection. Journal of Intelligent & Fuzzy https://www.geeksforgeeks.org/. Retrieved from
Systems, 1-9. geeksforgeeks.org:
[6]. Basha, S. H. (2022). An information-rich sampling https://www.geeksforgeeks.org/cnn-introduction-to-
technique over spatio-temporal CNN for classification pooling-layer/
of human actions in videos. Multimedia tools and [19]. Gobbert, M. K. (2021, April 3). High Performance
applications, 40431-40449. Computing Facility. Retrieved from hpcf.umbc.edu:
https://hpcf.umbc.edu/ada/slurm/overview/

IJISRT25FEB564 www.ijisrt.com 1224

Volume 10, Issue 2, February – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.5281/zenodo.14959415

[20]. Haohan Ding, J. T. (2023). The Application of https://www.scip.org/news/662235/Real-time-Threat-

Artificial Intelligence and Big Data in the Food Detection-The-Power-of-AI-in-Video-
Industry. MDPI, 29. Surveillance.htm
[21]. Heller, M. (2022, Sep 16). infoworld. Retrieved from [35]. Singh, S. (2023, July 20). What is a Conceptual
infoworld.com: Framework and How to Make It (with Examples).
https://www.infoworld.com/article/3299703/what-is- Retrieved from https://researcher.life/:
cuda-parallel-programming-for-gpus.html https://researcher.life/blog/article/what-is-a-
[22]. ibm.com. (2023, Dec 19). www.ibm.com. Retrieved conceptual-framework-and-how-to-make-it-with-
from www.ibm.com: examples/
https://www.ibm.com/think/topics/machine-learning- [36]. Stratified Random Sampling. (n.d.). Retrieved from
for-anomaly-detection Questionpro:
[23]. Kim, J. S. (2021). A study on implementation of real- https://www.questionpro.com/blog/stratified-random-
time intelligent video surveillance system based on sampling/
embedded module. EURASIP Journal on Image and [37]. Strickland, J. (2024, march 6). Parallel Computing.
Video Processing, 2-22. Retrieved from https://computer.howstuffworks.com/:
[24]. Kim, Y. M.-O. (2023). A Survey of Video https://computer.howstuffworks.com/parallel-
Surveillance Systems in Smart City. MDPI, 3-7. processing.htm
[25]. Kraus, J. (2013, March 13). An Introduction to CUDA- [38]. Techlab, N. (2024, June 12). www.netlabindia.com.
Aware MPI. Retrieved from Retrieved from www.netlabindia.com:
https://developer.nvidia.com/: https://www.netlabindia.com/blogs/emerging-
https://developer.nvidia.com/blog/introduction-cuda- technologies-shaping-the-future-of-video-
aware-mpi/ surveillance/
[26]. Mishra, M. (2020, Aug 26). Convolutional Neural [39]. Unzueta, D. (2022, October 18). Fully Connected
Networks, Explained. Retrieved from Layer vs. Convolutional Layer: Explained. Retrieved
towardsdatascience.com: from https://builtin.com/:
https://towardsdatascience.com/convolutional-neural- https://builtin.com/machine-learning/fully-connected-
networks-explained-9cc5188c4939 layer
[27]. Muts, I. (2023). Real-Time Video Monitoring: The [40]. Venujkvenk. (2023, Oct 03). Anomaly Detection
New Era of Remote Video Surveillance. Euristiq, 3- Techniques: A Comprehensive Guide with Supervised
14. and Unsupervised Learning. Retrieved from
[28]. Nielsen, F. (2016). Introduction to HPC with MPI for www.medium.com:
Data Science. Introduction to MPI: The Message https://medium.com/@venujkvenk/anomaly-
Passing Interface, 21-62. detection-techniques-a-comprehensive-guide-with-
[29]. Nishihara, R. (2019, Feb 11). Modern Parallel and supervised-and-unsupervised-learning-67671cdc9680
Distributed Python: A Quick Tutorial on Ray. [41]. Xidong Wu, P. B. (2023). Performance and Energy
Retrieved from https://towardsdatascience.com/: Consumption of Parallel Machine Learning
https://towardsdatascience.com/modern-parallel-and- Algorithms. ECE 2166.
distributed-python-a-quick-tutorial-on-ray- [42]. Yasar, K. (2023, March 31).
99f8d70369b8 https://www.techtarget.com/searchenterpriseai/defini
[30]. NVIDIA Tesla P100 PCIe 16 GB. (n.d.). Retrieved tion/image-recognition. Retrieved from
from techpowerup: www.techtarget.com:
https://www.techpowerup.com/gpu-specs/tesla-p100- https://www.techtarget.com/searchenterpriseai/definit
pcie-16-gb.c2888 ion/image-recognition
[31]. Rosebrock, A. (2021, May 14). Convolutional Neural [43]. Yim, J. (2015). Image Classification Using
Networks (CNNs) and Layer Types. Retrieved from Convolutional Neural Networks With Multi-stage
pyimagesearch.com: Feature. Robot Intelligence Technology and
https://pyimagesearch.com/2021/05/14/convolutional Applications 3, 587-594.
-neural-networks-cnns-and-layer-types/ [44]. Yokim, S. (2018, Aug 20). Tensor Ops Made Easier
[32]. Rosebrock, A. (2021, May 14). pyimagesearch. in cuDNN. Retrieved from developer.nvidia.com:
Retrieved from pyimagesearch.com: https://developer.nvidia.com/blog/tensor-ops-made-
https://pyimagesearch.com/2021/05/14/convolutional easier-in-cudnn/
-neural-networks-cnns-and-layer-types/ [45]. Zamora, R. R. (2021). Parallel and High Performance
[33]. Saha, S. (2018, Dec 15). A Comprehensive Guide to Computing. Parallel Programming (MPI) and Batch
Convolutional Neural Networks — the ELI5 way. Usage (SLURM), 650-704.
Retrieved from towardsdatascience.com:
https://towardsdatascience.com/a-comprehensive-
guide-to-convolutional-neural-networks-the-eli5-
way-3bd2b1164a53
[34]. Santilli, P. (2024, January 11). Real-time Threat
Detection: The Power of AI in Video Surveillance.
Retrieved from scip.org:

IJISRT25FEB564 www.ijisrt.com 1225