Ranjan.G, S. Akshatha, Sandeep.N and Vasanth.A, Acharya Institute of Technology, India

Computer Science & Engineering: An International Journal (CSEIJ), Vol 15, No 1, February 2025
DOI:10.5121/cseij.2025.15116 139
ENHANCING SURVEILLANCE SYSTEM
THROUGH EDGE COMPUTING: A
FRAMEWORK FOR REAL-TIME
HUMAN DETECTION
Ranjan.G. , Akshatha.S , Sandeep.N , Vasanth.A
Department of Computer Science and Engineering, Acharya Institute of
Technology, Bangalore, India
ABSTRACT
Real-time human detection is critical for modern surveillance, enabling timely responses to
security threats and enhancing situational awareness. Traditional approaches often
struggle with latency, bandwidth constraints, and centralized processing challenges. In this
paper, we propose a framework for real-time human detection using edge computing,
leveraging edge devices to process data closer to the source. Our framework employs
machine learning algorithms on edge devices to detect humans in real-time, reducing
latency and bandwidth usage. We detail the design and implementation of our framework,
including the edge computing architecture, machine learning models, and communication
protocols. Experimental results demonstrate the effectiveness and efficiency of the proposed
framework in real world scenarios underscoring its potential to enhance surveillance
systems across various applications.
KEYWORDS
Edge Computing, YOLO-V8, Surveillance System, Bandwidth, Latency.
1. INTRODUCTION
Surveillance systems play a crucial role in maintaining security and situational awareness in
various environments, ranging from public spaces and transportation hubs to private properties
and industrial facilities. Traditional surveillance systems often rely on centralized processing and
cloud-based analytics, which can lead to significant latency, bandwidth constraints, and potential
privacy concerns. These challenges hinder the ability to respond promptly to security threats and
limit the scalability and efficiency of surveillance operations.
In recent years, edge computing has emerged as a promising solution to these challenges by
bringing data processing closer to the data source. By leveraging the computational power of
edge devices, it is possible to perform real-time data analysis and decision-making at the edge of
the network. This decentralized approach reduces the need for constant data transmission to a
central server, thereby minimizing latency and bandwidth usage while enhancing data privacy
and system reliability.
In this paper, we propose a novel framework for real-time human detection using edge
computing, employing the latest YOLOv8 (You Only Look Once version 8) model. YOLOv8 is
renowned for its high detection accuracy and fast inference speed, making it particularly well

140
suited for real-time applications in resource-constrained environments. By deploying YOLOv8
on edge devices, our framework aims to achieve efficient and reliable human detection, enabling
timely responses to potential security threats.
The key contributions of this paper are as follows:
Edge Computing Framework: We design and implement an edge computing architecture that
integrates multiple edge devices to perform local processing and real-time human detection.
Model Deployment: We adapt and optimize the YOLOv8 model for deployment on edge devices,
ensuring high performance despite high performance despite limited computational resources.
Experimental Validation: We conduct extensive experiments in diverse real-world scenarios to
evaluate the effectiveness and efficiency of the proposed framework, demonstrating significant
improvements in latency bandwidth usage and scalability compared to traditional centralized
approaches.
2. REVIEW OF RELATED LITERATURE
2.1. Traditional Surveillance Systems
Traditional surveillance systems predominantly rely on centralized processing architectures
where data captured by cameras is transmitted to the central server for analysis. These systems
often face challenges related to latency and bandwidth constraints due to the transmission of
large volumes of raw data. Additionally, the centralized nature of these systems makes them
vulnerable to single points of failure, reducing their overall reliability and robustness in dynamic
environments.
2.2. Advances in Real-Time Human Detection
Real-time human detection has become a focal point of modern surveillance research due to its
critical role in ensuring timely responses to security threats. Various techniques have been
employed, including background subtraction, motion detection, and the use of advanced machine
learning algorithms. Notably, the introduction of convolution neural networks (CNNs) has
significantly improved detection accuracy. Models such as YOLO (You only Look Once and its
variants (e.g., YOLOv8, YOLOv4) have demonstrated high efficiency and accuracy in object
detection tasks, making them suitable for real-time human detection applications.
2.3. Edge Computing in Surveillance
Edge computing has emerged as a transformative paradigm in surveillance, addressing the
limitations of traditional centralized systems. By processing data closer to the source, edge
computing reduces latency and bandwidth usage, leading to faster response times and improved
system efficiency. Several studies have explored the integration of edge computing with
surveillance systems. For instance, Anwar et al. (2019) proposed an edge-based framework for
real-time object detection, highlighting significant reductions in latency and bandwidth
consumption.
Similarly, highlighting significant reductions in latency and bandwidth consumption . Similarly
Zang et al.(2020) demonstrated the benefits of deploying deep learning models on edge devices
for real-time video analytics in smart cities.

141
2.4. Machine Learning on Edge Devices
Deploying machine learning models on edge devices presents unique challenges, including
limited computational resources and power constraints. Recent advancements in model
optimization techniques, such as model pruning, quantization, and the development of
lightweight models like MobileNets, have enabled the effective deployment of deep learning
models on edge devices. These techniques reduce the computational load and energy
consumption while maintaining high accuracy, making them well-suited for real-time human
detection in surveillance systems.
2.5. Privacy and Security Considerations
The use of edge computing in surveillance also brings about critical privacy and security benefits.
By processing data locally on edge devices, sensitive information does not need to be transmitted
over the network, thereby reducing the risk of data breaches and unauthorized access. Studies by
Li et al. (2018) and Kim et al. (2019) have emphasized the importance of local data processing in
enhancing the privacy and security of surveillance systems. These studies highlight that edge
computing not only improves performance but also provides a more secure framework for
managing sensitive surveillance data.
2.6. Current Gaps and Opportunities
While significant progress has been made in integrating edge computing with surveillance
systems, there remain gaps and opportunities for further research. Most existing studies focus on
specific components of the system, such as the deployment of machine learning models or the
architectural design of edge networks. Comprehensive frameworks that address the end-to-end
integration of edge computing in real-time human detection are still relatively scarce.
Additionally, the evaluation of such systems in diverse real-world scenarios is needed to validate
their effectiveness and robustness.
3. METHODOLOGY
The proposed framework for real-time human detection through edge computing involves data
collection using a laptop camera, edge-based processing on the laptop, and centralized event
management through HTTP communication protocols and a web-based interface for viewing
playbacks and analytics. This section provides a detailed explanation of each component and the
overall system architecture as shown in Fig.3. In this paper, we utilize a laptop's built-in camera
to capture video streams continuously. The camera is configured to capture frames at a
predefined rate to ensure consistent data input for processing. Serving dual purposes, the laptop
also functions as the edge device, where all data processing tasks are executed. This setup allows
us to leverage the laptop's computational resources to perform real-time human detection using
the YOLOv8 model, processing the video frames locally to minimize latency and reduce the need
for extensive data transmission. By integrating data capture and processing within a single
device, we streamline the system architecture and enhance the overall efficiency and
responsiveness of the human detection framework.

142
Fig1. Edge Computing Architecture
In the edge-based processing phase, the raw video feed from the laptop's built-in camera
undergoes preprocessing to enhance the quality of the input data and reduce noise. This
preprocessing includes frame extraction, resizing, and normalization, preparing the video frames
for more effective analysis. The YOLOv8 (You Only Look Once version 8) model is selected for
human detection due to its high accuracy and real-time performance capabilities. This model is
deployed and run locally on the laptop, with optimizations to ensure it operates efficiently within
the constraints of the laptop’s computational resources.
The preprocessed video frames are then fed into the YOLOv8 model, which processes them to
detect human presence. The model outputs bounding boxes and confidence scores for each
detection. When a human is detected in the frame, the system initiates video recording. The
recording continues as long as a human is present in the frame. Once the model no longer detects
a human in the frame, the recording stops. The recorded video segment is then automatically
uploaded to the cloud for storage and further analysis. This approach ensures that only relevant
footage is recorded and transmitted, optimizing storage and bandwidth usage while enabling
efficient monitoring and timely response to security incidents. he centralized event management
system leverages Amazon Web Services (AWS) as the central server platform. AWS receives
and aggregates data packets from the laptop, ensuring efficient data handling and scalability.
Detected events are stored in an AWS-managed database.
A web-based interface is developed to provide users with convenient access to recorded
playbacks and analytics.
This interface features playback functionality, allowing users to view recorded video segments
where human detection events occurred. During playback, bounding boxes are displayed around
detected individuals to highlight their presence in the frame. Additionally, the interface includes
an analytics dashboard that presents statistical data and visualizations related to detection events.
This dashboard provides insights into detection patterns, frequency, and other relevant metrics,
enabling users to perform detailed analysis and derive actionable insights from the surveillance
data.

143
Fig.2. Methodology flowchart
Fig.3. Sequence diagram
The YOLOv8 (You Only Look Once version 8) model is selected for this project due to its high
accuracy and real-time performance capabilities, which are crucial for effective human detection
in surveillance systems. YOLOv8 is part of the YOLO family of models, renowned for their
ability to perform object detection swiftly and accurately in a single stage. Unlike traditional
object detection methods that rely on a two-stage approach—first generating region proposals
and then classifying them—YOLOv8 performs detection in a single stage. This approach
significantly enhances the model’s speed, making it highly suitable for applications requiring
real-time processing.
The YOLOv8 algorithm works by dividing the input image into a grid, with each grid cell being
responsible for detecting objects whose centers fall within that cell. Specifically, the input image

144
is divided into an S×S grid as shown in Fig.4. Each grid cell predicts a fixed number of bounding
boxes, with each prediction including the coordinates of the box’s center relative to the grid cell,
the width and height of the box relative to the entire image, a confidence score indicating the
likelihood of an object being present, and the class probabilities indicating which class the
detected object belongs to (e.g., human, car, dog). This grid-based approach allows YOLOv8 to
detect multiple objects in close proximity efficiently.
Fig.4. Image divided into a SxS grid
To improve detection accuracy and remove redundant bounding boxes, YOLOv8 employs Non-
Maximum Suppression (NMS) as shown in Fig.5. NMS is a technique that eliminates
overlapping bounding boxes, keeping only the one with the highest confidence score for each
detected object. This ensures that each human detected in the video frame is represented by a
single, accurate bounding box, enhancing the reliability of the detection results.
The feature extraction process in YOLOv8 involves a deep convolutional neural network (CNN)
that progressively extracts higher-level features from the input image through multiple
convolutional layers. These layers capture essential details needed for accurate object detection.
During training, YOLOv8 uses a multi-part loss function that combines localization loss (to
measure the accuracy of the predicted bounding box coordinates), confidence loss (to measure
the accuracy of the predicted confidence scores), and class probability loss (to measure the
accuracy of the predicted class probabilities).
Fig.5.YOLOv8 employing Non Maximum Suppression (NMS)
In this project, YOLOv8 is fine-tuned with specific data relevant to the deployment environment
to ensure precise human detection under various conditions. The model is deployed on a laptop,
which serves as the edge device, and is optimized to run efficiently within the laptop's
computational constraints to enable real-time inference.

145
4. IMPLEMENTATION
The implementation of our human detection system begins with the hardware setup, where a
laptop with a built-in camera serves as the edge device. This laptop, equipped with an Intel i5
processor, 8GB of RAM, and an integrated GPU, captures video at 1080p resolution. The
software environment is built on Fedora, utilizing Python as the primary programming language.
Key libraries and frameworks include YOLOv8n for object detection, OpenCV for video capture
and preprocessing, PyTorch for running the YOLOv8n model, AWS SDK (Boto3) for cloud
integration, and ReactJs for developing the web-based interface and analytics dashboard.
The data collection process involves configuring the laptop’s built-in camera to continuously
capture video streams. OpenCV is utilized to capture video frames, which are then preprocessed
to enhance the quality and reduce noise. This preprocessing includes frame extraction, resizing to
640x640 pixels, and normalization to ensure consistency in the input data.
The YOLOv8n model, represented by the yolov8n.pt file, is selected for its lightweight nature
and efficient performance. This pre-trained model is capable of detecting multiple object types,
but it is fine-tuned with a focus on detecting humans. PyTorch is used to load the YOLOv8n
model, which is then deployed locally on the laptop for real-time inference. The model is
optimized to run efficiently within the computational constraints of the edge device, ensuring
rapid detection of humans in video streams.
To ensure that only human detections are considered, a post-processing filter is applied to the
output of the YOLOv8n model. This filter identifies and retains only those detections labeled as
‘person’, discarding detections related to other object types. By filtering the model’s output, the
system focuses exclusively on detecting humans, which is essential for the intended surveillance
application.
Fig.6.Consoling the person detected
The real-time detection process begins with the preprocessed video frames being fed into the
YOLOv8n model. The model analyzes each frame to detect the presence of humans, generating
bounding boxes around detected individuals. If humans are detected as shown in Fig.7 in the
frame, the system initiates video recording. The recording continues as long as humans are
present in the video stream and stops when no humans are detected. This adaptive recording
mechanism ensures that only relevant footage is captured, optimizing storage and bandwidth
usage.

146
Fig.7. YOLOv8 performing human detection
AWS serves as the central server for aggregating data packets from the edge device and
managing the storage of recorded video segments. Boto3, the AWS SDK for Python, is used to
upload recorded videos to AWS S3 for secure and scalable storage. By leveraging AWS services,
the system ensures reliable data storage and seamless integration with cloud-based infrastructure.
A web-based interface developed using ReactJs provides users with access to recorded playbacks
as shown in Fig.8. and an analytics dashboard. The playback functionality allows users to view
recorded video segments with bounding boxes displayed around detected humans, facilitating
visual verification of detection results. The analytics dashboard presents statistical data and
visualizations related to detection events, offering insights into detection patterns and frequency.
This intuitive interface enhances user experience and enables efficient monitoring of surveillance
data.
Fig.8. Website playback page

147
Fig.9. Video playback
The system undergoes comprehensive testing to ensure that each component functions correctly
and meets performance expectations. Testing procedures include unit testing, integration testing,
and system testing to validate the functionality and reliability of the entire system. Key
performance metrics such as detection accuracy, processing latency, and system throughput are
evaluated to assess the system’s effectiveness in real-world scenarios.
Deployment involves setting up the edge device and configuring AWS services according to the
system requirements. Detailed deployment instructions are provided to facilitate the setup
process. Ongoing maintenance procedures are established for regular updates and model
retraining to ensure that the system remains efficient and up-to-date with evolving requirements
and technologies.
5. RESULTS
The system's effectiveness in real-time human detection was demonstrated through various tests,
where it accurately detected and tracked humans in dynamically changing environments. The
system's ability to start and stop recording based on human presence ensures efficient use of
storage and bandwidth. Example frames from these tests showed bounding boxes accurately
drawn around detected humans, illustrating the system's robustness in different lighting
conditions and background complexities.
Additionally, we implemented an analytics graph to visualize the number of persons detected
over different time periods, providing valuable insights into patterns of human activity. The daily
detection graph as shown in Fig.10. showed the number of persons detected each hour,
highlighting peaks of high activity and informing resource allocation during busy periods. The
monthly detection graph aggregated daily detections, revealing trends over the month and
identifying days with unusually high or low activity. The yearly detection graph provided an
overview of detections throughout the year, showing seasonal variations and long-term trends.
These analytics graphs offer a comprehensive overview of human detection patterns, aiding in
the optimization of surveillance operations and security measures.
Fig.10. Daily Analytics

148
Throughout the implementation, we faced challenges related to hardware limitations and
environmental variability. For instance, in low-light conditions, the detection accuracy slightly
decreased. However, these challenges were mitigated by preprocessing techniques and careful
tuning of the model parameters.
6. CONCLUSION AND FUTURE RECOMMENDATIONS
In conclusion, the implementation of our human detection system has yielded promising results
in enhancing surveillance capabilities for security applications. Through the utilization of the
YOLOv8n model and edge computing, we have successfully developed a system capable of real-
time detection of humans in video streams captured by a laptop's built-in camera. The system
demonstrates commendable performance metrics, including high detection accuracy and minimal
processing latency, thus enabling timely response to security threats and enhancing situational
awareness. By filtering the YOLOv8n model's output to focus exclusively on human detections,
we have effectively addressed the specific requirements of our surveillance application.
However, it's important to acknowledge the limitations encountered during implementation,
including constraints related to hardware resources and environmental variability. Despite these
challenges, the human detection system showcases significant potential for improving security
operations and facilitating proactive threat mitigation strategies.
Looking ahead, there are several avenues for enhancing the capabilities and effectiveness of the
human detection system. Firstly, further refinement and optimization of the YOLOv8n model
could be pursued to improve detection accuracy and efficiency. This may involve fine-tuning
model parameters, incorporating additional training data, or exploring advanced architectural
modifications. Additionally, the integration of advanced computer vision techniques, such as
multi-object tracking and pose estimation, could enrich the system's understanding of human
behavior and improve contextual awareness. Upgrading the hardware infrastructure to support
larger-scale deployments and exploring the integration of sensor networks for complementary
sensory information are also viable avenues for future enhancements. Furthermore,
enhancements to the user interface and interaction design could improve usability and
accessibility for security personnel, while extensive field testing and real-world deployment will
provide valuable insights for validating the system's performance and identifying areas for further
refinement. Overall, these future enhancements aim to propel the human detection system
towards greater effectiveness and applicability in diverse surveillance scenarios, ultimately
contributing to the advancement of security technology and situational awareness capabilities.
REFERENCES
[1] Sayed Yahya Nikouei, Yu Chen, Sejun Song, Rongua Xu, Baek-Young Choi, Timothy R.
Faughnan, Smart surveillance as an edge network service: from Haar-cascade, SVM to a
Lightweight CNN, 2018
[2] Rajkumar Rajavel, SathishKumar. Ravichandran, Karthikeyan Harimoorthy, Partheeban
Nagappan, IoT-Based smart healthcare video surveillance system using edge computing, J.
Ambient Intel. Humaniz. Comput. (2021).
[3] K. N. Karthick Kumar, H. Ntraj, and T. P. Jacob, “Motion activated security camera using
Raspberry Pi,” International conference on Communication and Signal Processing (ICCSP),2017.
[4] W. Yang and Z. Jiachun, “Real-time face detection based on YOLO,” International
conference on Knowledge Innovation and Invention, 2018.
[5] Edge computing based surveillance framework for real time activity recognition Aishwarya D.∗,
Minu R.I., 2021
[6] Chulyeon Kim, Jiyoung Lee, Taekjin Han, Young-Min Kim, A hybrid
framework combining background subtraction and deep neural network for rapid person
detection, J. Big Data 5 (2018) 22.

149
[7] M. Satyanarayanan. 2017. The Emergence of Edge Computing. Computer 50, 1(Jan 2017), 30–39.
Generation Healthcare. Studies in Big Data, C. Bhatt, N. Dey, and A. Ashour, Eds., vol. 23,
Springer, Cham., 2017.
[8] Jianyu Wang, Jianli Pan, and Flavio Esposito. 2017. Elastic Urban Video Surveillance System
Using Edge Computing. In Proceedings of SmartIoT’17, San Jose / Silicon Valley, CA, USA,
October 14, 2017
[9] Video surveillance systems-current status and future trends. Vassilios Tsakanikas ∗, Tasos
Dagiuklas (2017)
[10] Combining Background Subtraction and Convolutional Neural Network for Anomaly Detection in
Pumping-Unit Surveillance. Tianming Yu ,Jianhua Yang and Wei Lu (2019)
[11] A Home security camera system with container based resource allocation on Raspberry pi , Takuya
Egashira, Hiroki Nishikawa, Xiangbo Kong, Hiroyuki Tomiyama(2021).

Ranjan.G, S. Akshatha, Sandeep.N and Vasanth.A, Acharya Institute of Technology, India

More Related Content

Similar to Ranjan.G, S. Akshatha, Sandeep.N and Vasanth.A, Acharya Institute of Technology, India (20)

More from CSEIJJournal (20)

Recently uploaded (20)

Ranjan.G, S. Akshatha, Sandeep.N and Vasanth.A, Acharya Institute of Technology, India