VISVESVARAYA TECHNOLOGICAL UNIVERSITY

JNANA SANGAMA, BELAGAVI - 590 018

A project Phase I Report on

Hawk-Eye, AI-Powered Smart Surveillance with
Anomaly Detection and Behaviour analysis
Submitted in partial fulfillment of the requirements for the VI Semester of degree
of Bachelor of Engineering in Information Science and Engineering of
Visvesvaraya Technological University, Belagavi

Submitted By

Pramod R Ilal 1RN22IS110 Nagaraj S Bhat 1RN22IS090

Mohit S Bidanur 1RN22IS087 Manoj K S 1RN22IS082

Under the Guidance of

Ms. Nivedita. G Y

Assistant Professor
Department of ISE

Department of Information Science and Engineering

RNS Institute of Technology
Dr. Vishnuvaradhan Road, Rajarajeshwari Nagar post,
Channasandra, Bengaluru-560098
2024-2025
 RNS INSTITUTE OF TECHNOLOGY
Dr. Vishnuvaradhan Road, Rajarajeshwari Nagar post,
Channasandra, Bengaluru - 560098
DEPARTMENT OF INFORMATION SCIENCE AND ENGINEERING

CERTIFICATE
Certified that the project work phase 1 entitled Hawk_Eye – AI powered smart surveillance
with anomaly detection and behaviour analysis has been successfully completed by Pramod
Raghavendra Ilal (1RN22IS110), Nagaraj Subraya Bhat (1RN22IS090), Mohit
Shamsundar Bidanur (1RN22IS087), Manoj K S (1RN22IS082) bona fide students of RNS
Institute of Technology, Bengaluru in partial fulfillment of the requirements for the award
of degree in Bachelor of Engineering in Information Science and Engineering of
Visvesvaraya Technological University, Belagavi during academic year 2024-2025. The
project phase 1 report has been approved as it satisfies the academic requirements in respect of
project phase 1 work for the said degree.

Ms.Nivedita G Y Dr. Suresh L Dr. Ramesh Babu

Project Guide Professor and HoD Principal
Assistant Professor Department of ISE RNSIT
Department of ISE RNSIT

External Viva
Name of the Examiners Signature with Date

1. _ 1. _ _

_
2. _ 2.
 DECLARATION

We, Pramod Raghavendra Ilal [1RN22IS110], Nagaraj Subraya Bhat [1RN22IS090], Mohit
Shamsundar Bidanur [1RN22IS087] and Manoj K S [1RN22IS082], students of VI Semester
BE, in Information Science and Engineering, RNS Institute of Technology hereby declare that
the Project work phase 1 entitled Attendance Management System Using Face Recognition has
been carried out by us and submitted in partial fulfillment of the requirements for the VI Semester
degree of Bachelor of Engineering in Information Science and Engineering of Visvesvaraya
Technological University, Belgaum during academic year 2024-2025.

Place: Bengaluru

Date:

PRAMOD R ILAL[1RN22IS110]

NAGARAJ S BHAT [1RN22IS090]

MOHIT S BIDANUR [1RN22IS087]

MANOJ K S [1RN22IS082]
 ABSTRACT

In this project, we propose an intelligent surveillance system that combines anomaly

detection, unattended object monitoring, and behavioural analysis to enhance security in
monitored environments. Leveraging deep learning techniques and computer vision, the
system processes pre-recorded surveillance video data to identify unusual events and analyse
human activity patterns.

The anomaly detection module uses a convolutional autoencoder-based architecture to detect

deviations from learned normal motion patterns in video sequences. This enables early
detection of potentially suspicious behaviour without the need for predefined rules.

To further strengthen the surveillance capabilities, an unattended object detection feature is

incorporated. This module continuously tracks objects across video frames and raises alerts if
an object remains stationary and unaccompanied for a predefined period, indicating possible
security threats such as left baggage.

Additionally, the system performs basic behaviour analysis by interpreting abnormal activity
as a deviation from normal movement patterns, thus providing insights into suspicious or
unexpected behaviour without explicitly classifying individual actions. This behaviour
analysis contributes to a more holistic understanding of the surveillance context.

Overall, the integrated framework offers a robust, offline video surveillance solution capable
of highlighting anomalous and unattended events while supporting basic activity-based
behavioural understanding.

I
 ACKNOWLEDGMENT

The fulfillment and rapture that go with the fruitful finishing of any assignment would be
inadequate without the specifying the people who made it conceivable, whose steady direction and
support delegated the endeavors with success.

We would like to profoundly thank Management of RNS Institute of Technology for

providing such a healthy environment to carry out this project work.

We would like to express our thanks to our Principal Dr. Ramesh Babu for his support and
inspired us towards the attainment of knowledge.

We wish to place on record my words of gratitude to Dr. Suresh L, Professor and Head of the
Department, Information Science and Engineering, for being the enzyme and master mind behind our
project work.

We place our heartfelt thanks to Dr. Hema N, Associate Professor, Department of

Computer Science and Engineering for having guided the project and all the staff members of the
department of Information Science and Engineering for helping at all times.

We would also like to thank our project coordinator Ms. Nivedita G Y, Assistant Professor,
Department of ISE, RNSIT, Bangalore, for their valuable suggestions.

We would like to thank all other teaching and non-teaching staff of Information Science &
Engineering, RNSIT, Bangalore, who have directly or indirectly helped us to carry out the project
work

Place: Bengaluru
Date:
PRAMOD R ILAL[1RN22IS110]

NAGARAJ S BHAT [1RN22IS090]

MOHIT S BIDANUR [1RN22IS087]

MANOJ K S [1RN22IS082]

II
 TABLE OF CONTENTS

CERTIFICATE
DECLERATION
ABSTRACT I
ACKNOWLEDGMENT II
TABLE OF CONTENTS III
LIST OF FIGURES IV
ABBREVATIONS V
1. INTRODUCTION 1
2. LITERATURE SURVEY 3
3. PROBLEM STATMENT 17
4. OBJECTIVES 18
5. METHODOLOGY 19
6. SYSTEM DESIGN 25
7. CONCLUSION 29
8. REFERENCE 30

III
 LIST OF FIGURES
Fig No. Description Page No.
Figure 5.1 Structural Chart of the proposed system 19
Figure 6.1 Flowchart of the system 28

IV
LIST OF ABBREVIATIONS
AI – Artificial Intelligence

CNN – Convolutional Neural Network

RNN – Recurrent Neural Network

GAN – Generative Adversarial Network

YOLO – You Only Look Once

CAE – Convolutional Autoencoder

MSE – Mean Squared Error

UI – User Interface

LSTM – Long Short-Term Memory

SIEM – Security Information and Event Management

V
 Chapter 1
INTRODUCTION

1.1 Overview
In the modern world, video surveillance plays a vital role in ensuring public safety and
monitoring environments like airports, train stations, and public places. Traditional systems
depend heavily on human operators to detect suspicious behavior, which can be time-
consuming, error-prone, and inefficient when monitoring multiple video feeds. To address these
limitations, this project introduces an intelligent video surveillance system using machine
learning for anomaly detection, unattended object tracking, and behavior analysis.

The system is designed to process offline video footage and identify unusual events based on
patterns learned from normal behavior. Using deep learning (ConvLSTM Autoencoders), the
model detects frame-level anomalies without needing labeled abnormal data. If an activity
deviates from expected motion patterns, it is marked as abnormal.

Additionally, the project includes an unattended object detection feature. When an object is left
unattended for a certain time, the system flags it, which is useful in public safety applications.

Behavior analysis builds upon the anomaly detection module by interpreting detected
irregularities. While it does not explicitly label activities, it helps understand what kind of event
is occurring based on deviations in behavior, contributing to smarter surveillance decisions.

1.2 Project Description

This project presents a Smart Surveillance System that analyzes pre-recorded video data to
automatically detect abnormal activities, understand behavior patterns, and monitor unattended
objects. The system employs machine learning and computer vision techniques to process
surveillance footage, identify anomalies that deviate from learned normal behavior, and classify
basic actions occurring within the scene.
Anomaly detection is performed using a convolutional autoencoder-based deep learning model,
which learns the typical patterns from training video sequences. When new video frames are
analyzed, the model identifies significant deviations as potential anomalies—such as sudden
movements, running, or unexpected object appearances.The backend, powered by Node.js and
Express.js, ensures secure and efficient data.

Department of ISE, RNSIT 2024 – 2025 1

Hawk-Eye Smart Surveillance

The behavior analysis component attempts to categorize the nature of actions present in the
scene (e.g., walking, standing, bending) using spatial-temporal features extracted from the
videos. Additionally, the system includes an unattended object timer that flags objects left in
one location for too long, indicating possible security risks.
The project is implemented using Python, TensorFlow, OpenCV, and Streamlit for a user-
friendly interface. While real-time processing is not included, this system effectively
demonstrates the core concepts of intelligent video analysis through offline video datasets.
While real-time video processing is not implemented, this system effectively demonstrates the
core principles of intelligent video surveillance through offline video datasets. It simulates how
modern surveillance systems can enhance security and awareness in environments like
campuses, offices, and public areas using machine learning and video analytics.
The project is built using Python, TensorFlow, OpenCV, and Streamlit, offering a user-friendly
interface to visualize outputs such as abnormal event detection, identified behaviors, and object
tracking timelines. Users can upload pre-recorded videos and analyze them through the
application, gaining insights into various scene dynamics.
One of the system’s key features is the detection of unattended objects. The system continuously
monitors objects in the video and triggers a timer when an object is left unmoved for a specific
period, helping identify potential security threats such as abandoned luggage.

Department of ISE, RNSIT 2024 – 2025 2

 Chapter 2

LITERATURE REVIEW

A literature survey or a literature review in a project report shows the various

analyses and research made in the field of interest and the results already published,
taking into account the various parameters of the project and the extent of the project.
Literature survey is mainly carried out in order to analyze the background of the current
project which helps to find out flaws in the existing system & guides on which unsolved
problems we can work out. So, the following topics not only illustrate the background
of the project but also uncover the problems and flaws which motivated to propose
solutions and work on this project.

A literature survey is a text of a scholarly paper, which includes the current

knowledge including substantive findings, as well as theoretical and methodological
contributions to a particular topic. Literature reviews use secondary sources, and do not
report new or original experimental work. Most often associated with academic-oriented
literature, such as a thesis, dissertation or a peer-reviewed journal article, a literature
review usually precedes the methodology and results sectional though this is not always
the case. Literature reviews are also common in are search proposal or prospectus (the
document that is approved before a student formally begins a dissertation or thesis). A literature
survey includes the following:

• Existing theories about the topic which are accepted universally.

• Books written on the topic, both generic and specific.

Objectives of Literature Survey

• Learning the definitions of the concepts.

• Access to latest approaches, methods and theories.
• Discovering research topics based on the existing research

Department of ISE, RNSIT 2024 – 2025 3

Hawk-Eye Smart Surveillance

2.1. Title: "AI-powered surveillance systems and anomaly detection."

Author: Chris, E., Olabiyi, W., Favor, B., & Ramon, T.
Year:2023
Abstract: The authors explore the development and implementation of intelligent
surveillance systems powered by AI technologies such as machine learning, computer
vision, and neural networks. The paper focuses on how AI enables real-time analysis of
vast video data to detect suspicious behaviors and abnormal activities. Techniques like
Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are
employed to learn temporal and spatial features. It also emphasizes the integration of
IoT devices for data collection, edge computing for faster processing, and cloud storage
for scalability. The ethical implications, including data privacy and surveillance
overreach, are acknowledged, and solutions such as encryption, anonymization, and
policy regulation are discussed. Overall, the study highlights the transformational
potential of AI in proactive threat detection and decision-making in security systems.
Limitations: AI-powered surveillance systems face key limitations. Privacy remains a
major concern, as constant monitoring can infringe on individual rights and freedoms.
These systems may also reflect biases present in their training data, leading to unfair or
inaccurate detections. Additionally, they require high computational resources, making
them expensive and energy-intensive to operate.

2.2. Title: "Deep Learning-Based Anomaly Detection in Video Surveillance."

Author: Huu-Thanh Duong, Viet-Tuan Le, and Vinh Truong Hoang
Year:2023
Abstract: This study presents a structured overview of various anomaly detection
methods used in intelligent surveillance. It delves into deep learning approaches,
especially CNNs and autoencoders, for feature extraction and pattern recognition. The
paper describes supervised and unsupervised learning techniques, evaluating their pros
and cons for real time deployment. It also explains the use of anomaly scoring to rank
unusual behaviors and prioritize alerts. The importance of large-scale labeled datasets is
emphasized for training robust models. Training methodologies, evaluation metrics
(e.g., precision, recall, F1-score), and validation techniques are analyzed in detail.
Furthermore, the paper explores data augmentation strategies, real-time video analytics
frameworks, and deployment challenges like occlusions, poor lighting, and crowd

Department of ISE, RNSIT 2024 – 2025 4

Hawk-Eye Smart Surveillance

density. It concludes by advocating AI to enhance automation in surveillance and reduce

human fatigue and error in security operations.
Limitations: A significant challenge is the scarcity of diverse, labeled datasets, which
hampers the models' ability to generalize across various environments and scenarios.
Additionally, many models struggle with real-time processing due to high
computational demands, limiting their practical deployment. The lack of explainability
in deep learning models also poses issues for trust and interpretability, especially in
critical surveillance applications. Furthermore, these systems often exhibit sensitivity to
environmental variations, such as lighting changes or camera angles, leading to
decreased robustness in dynamic real-world settings.

2.3. Title: "AI-Powered Anomaly Detection Techniques"

Author: Mia Cate.
Year:2023
Abstract: This paper introduces a hybrid surveillance system architecture where edge
devices are utilized to perform real-time video processing and anomaly detection,
significantly reducing dependency on cloud servers. The authors integrate YOLO (You
Only Look Once) models with edge computing to ensure high-speed object detection
and classification. The performance is benchmarked using datasets such as Avenue and
ShanghaiTech, where the system achieves low latency and energy-efficient processing.
The discussion includes technical aspects like model compression, edge resource
management, and balancing computational cost with detection accuracy. It explores
benefits such as reduced data transmission, enhanced privacy, and faster response times.
Additionally, the system’s ability to scale across multiple surveillance points is assessed,
along with its adaptability to dynamic environments. Limitations, including memory
constraints of edge devices and challenges in model updates, are critically analyzed.
Limitations: One major issue is the difficulty in obtaining high-quality, labeled datasets
for training, which impacts model accuracy. Many AI models struggle with real-time
processing, making them less effective in time-sensitive applications. Additionally,
explainability remains a concern, as black-box models often provide little insight into
their decision-making process, reducing trust and transparency. The presence of noisy
or imbalanced data further complicates anomaly detection, leading to false positives or
negatives.

Department of ISE, RNSIT 2024 – 2025 5

Hawk-Eye Smart Surveillance

2.4. Title: "Revolutionizing Video Surveillance: AI-Powered Anomaly Detection."

Author: Nishant Deheriya, Devendra Kumar Bajpai, and P.K. Sharma.
Year:2023
Abstract: The paper evaluates several anomaly detection techniques including
statistical, machine learning, and hybrid models. It examines algorithms such as k-NN,
Isolation Forest, One Class SVM, and autoencoders, comparing them based on metrics
like accuracy, precision, recall, computational complexity, and latency. A testbed using
synthetic and real-world surveillance datasets is created to measure performance under
various scenarios including low light, crowd behavior, and rapid motion. The authors
propose ensemble models that combine the strengths of different algorithms to improve
robustness and adaptability. Furthermore, system architecture recommendations are
provided for integrating these algorithms into surveillance platforms. Discussions also
include how models handle false positives and dynamic environments, making this a
comprehensive resource for developers choosing the right anomaly detection strategy.
Limitations: These include challenges related to the accuracy of detection algorithms,
especially in complex environments, and the potential for false positives or negatives.
Additionally, the deployment of such AI systems raises concerns about privacy and the
ethical implications of constant surveillance. The authors suggest that addressing these
issues is crucial for the effective and responsible implementation of AI-powered
surveillance systems.

2.5. Title: "AI-powered threat detection in surveillance systems: A real-time data

processing framework."
Author: Emmanuel Cadet, Olajide Soji Osundare, Harrison Ekpobimi, Zein Samira,
and Yodit Wondaferew Weldegeorgise.
Year:2024
Abstract: The research highlights a vision-based AI framework designed to monitor,
analyze, and identify anomalous human behaviors through real-time video feeds. It
utilizes deep learning models like CNNs for spatial feature extraction, RNNs for
temporal analysis, and Generative Adversarial Networks (GANs) to generate and
evaluate synthetic anomaly samples. The system architecture includes cloud

Department of ISE, RNSIT 2024 – 2025 6

Hawk-Eye Smart Surveillance

connectivity for centralized management, edge analytics for instant alert generation, and
an intuitive dashboard for operators. Specific use cases are discussed such as unattended
baggage detection, loitering, and crowd congestion. The authors emphasize real-time
alerting, scalability, and seamless integration with existing CCTV systems. Moreover,
they address key challenges including data imbalance, interpretability of AI decisions,
and hardware-software compatibility. The proposed solution shows promise for urban
security, airport monitoring, and public event management.
Limitations: While the paper presents a comprehensive framework for real-time AI-
powered threat detection in surveillance systems, it also acknowledges certain
limitations. These include challenges related to the accuracy of detection algorithms,
especially in complex environments, and the potential for false positives or negatives.
Additionally, the deployment of such AI systems raises concerns about privacy and the
ethical implications of constant surveillance. The authors suggest that addressing these
issues is crucial for the effective and responsible implementation of AI-powered
surveillance systems.

2.6. Title: "Towards a Conceptual Framework for AI-Driven Anomaly Detection in Smart
City IoT Networks"
Author: Heng Zeng , Manal Yunis, Ayman Khalil
Year:2024
Abstract: This paper presents a conceptual framework designed to detect anomalies and
security threats in smart city environments using AI techniques. It discusses the unique
challenges in IoT networks, such as data heterogeneity, real-time response needs, and
complex sensor interconnectivity. The proposed framework integrates machine learning
with context-aware anomaly detection, focusing on scalability and interoperability. It
includes data preprocessing, feature extraction from IoT devices, and adaptive learning
modules. The model emphasizes flexibility to accommodate different urban
infrastructures and threat types. The authors suggest integrating edge and cloud
computing for real-time processing, and outline potential use cases like traffic
monitoring, utility management, and emergency response. Future research directions
include addressing privacy concerns and improving explainability in AI decision
making.
Limitations: The paper identifies several challenges in implementing AI-driven
anomaly detection within smart city IoT networks. One significant issue is the difficulty

Department of ISE, RNSIT 2024 – 2025 7

Hawk-Eye Smart Surveillance

in obtaining high-quality, labeled datasets for training, which impacts model accuracy.
Many AI models struggle with real-time processing, making them less effective in time-
sensitive applications. Additionally, explainability remains a concern, as black-box
models often provide little insight into their decision-making process, reducing trust and
transparency. The presence of noisy or imbalanced data further complicates anomaly
detection, leading to false positives or negatives.

2.7. Title: "AI-Based Anomaly Detection for Real-Time Cybersecurity"

Author: Maloy Jyoti Goswami.
Year:2023
Abstract: This paper examines the role of AI in enabling real-time anomaly detection
for cybersecurity. It outlines a modular AI architecture that combines supervised and
unsupervised learning techniques for timely identification of cyber threats. Key
components include a streaming data pipeline, a behaviour profiling engine, and an alert
prioritization system. Algorithms such as Isolation Forest, Support Vector Machines
(SVM), and deep neural networks are evaluated. The system is tested on datasets such
as NSL-KDD and CICIDS2017, showing improvements in both detection accuracy and
latency. The authors emphasize low false positive rates and the ability to adapt to novel
attack vectors. Integration with Security Information and Event Management (SIEM)
systems is discussed to enhance operational response.
Limitations: The paper identifies several challenges in implementing AI-based
anomaly detection within real-time cybersecurity systems. One significant issue is the
need for large, high-quality datasets to train machine learning models effectively, which
can be resource-intensive to obtain. Additionally, the computational overhead
associated with processing real-time data streams can strain system resources. The risk
of adversarial attacks targeting AI models is another concern, as malicious actors may
attempt to deceive the system. The paper discusses potential mitigation strategies for
these challenges, emphasizing the importance of robust data handling and model
validation techniques.

2.8. Title: "Deep Learning-Based Anomaly Detection in Video Surveillance: A Survey "
Author: Huu-Thanh Duong Viet-Tuan Le Vinh Truong Hoang.
Year:2023

Department of ISE, RNSIT 2024 – 2025 8

Hawk-Eye Smart Surveillance

Abstract: This survey reviews the landscape of deep learning-based anomaly detection
in video surveillance. It classifies methods based on architecture—CNNs, RNNs,
autoencoders, GANs, and hybrid models. The paper compares various algorithms in
terms of accuracy, processing speed, and scalability. It discusses benchmark datasets
such as UCSD Ped1/Ped2, Shanghai Tech, and UCF-Crime. Key challenges identified
include scarcity of labelled anomaly data, high computational cost, and the
generalization gap between training and deployment environments. The survey
highlights trends toward using self-supervised learning, attention mechanisms, and
transformer models. Future directions include real-time edge computing, privacy-
preserving analytics, and cross-domain anomaly detection.
Limitations: The paper identifies several challenges in implementing deep learning-
based anomaly detection within video surveillance systems. One significant issue is the
difficulty in obtaining high-quality, labeled datasets for training, which impacts model
accuracy. Many AI models struggle with real-time processing, making them less
effective in time-sensitive applications. Additionally, explainability remains a concern,
as black-box models often provide little insight into their decision-making process,
reducing trust and transparency. The presence of noisy or imbalanced data further
complicates anomaly detection, leading to false positives or negatives.

2.9. Title: "AI-Powered Anomaly Detection for Real-Time Performance Monitoring in

Cloud Systems."
Author: Pushpalika Chatterjee, Apurba Das.
Year:2023
Abstract: The paper focuses on anomaly detection in cloud computing environments,
highlighting the shortcomings of traditional rule-based and statistical methods. It
proposes a unified AI powered model that combines system metrics monitoring, time-
series analysis, and anomaly pattern recognition. The model uses recurrent neural
networks (RNNs) and autoencoders to capture both short- and long-term performance
trends. Datasets from public cloud monitoring tools (e.g., AWS CloudWatch) are used
for validation. The system demonstrates improved detection rates for CPU spikes,
memory leaks, and latency anomalies. The authors propose dynamic model updates
through online learning and suggest integrating explainable AI (XAI) techniques to
enhance trust and traceability in anomaly reports.

Department of ISE, RNSIT 2024 – 2025 9

Hawk-Eye Smart Surveillance

Limitations: The paper identifies several challenges in implementing AI-powered

anomaly detection within cloud systems. One significant issue is the difficulty in
obtaining high-quality, labeled datasets for training, which impacts model accuracy.
Many AI models struggle with real-time processing, making them less effective in time-
sensitive applications. Additionally, explainability remains a concern, as black-box
models often provide little insight into their decision-making process, reducing trust and
transparency. The presence of noisy or imbalanced data further complicates anomaly
detection, leading to false positives or negatives.

2.10. Title: "Video Anomaly Detection System Using Deep Convolutional and Recurrent
Neural Networks."
Author: Maryam Qasim, Elena Verdu.
Year:2023
Abstract: This paper presents a deep learning model that integrates Convolutional
Neural Networks (CNNs) for spatial feature extraction and Recurrent Neural Networks
(RNNs), particularly LSTM units, for temporal pattern recognition in video anomaly
detection. The model is designed to detect unusual behaviours such as running in
restricted areas or loitering. A two stage pipeline processes raw video frames into spatial
features and sequences, which are then analysed for deviations from learned normal
patterns. Evaluation on datasets like CUHK Avenue and Shanghai Tech shows F1-
scores above 90%. The authors address challenges like temporal coherence and real-
time inference by optimizing the model using pruning and frame sampling. The hybrid
approach achieves robust performance across diverse video scenarios, suggesting
applicability in both public and private surveillance systems.
Limitations: The paper identifies several challenges in implementing deep
convolutional and recurrent models for video anomaly detection. One significant issue
is the difficulty in obtaining high-quality, labeled datasets for training, which impacts
model accuracy. Many AI models struggle with real-time processing, making them less
effective in time-sensitive applications. Additionally, explainability remains a concern,
as black-box models often provide little insight into their decision-making process,
reducing trust and transparency. The presence of noisy or imbalanced data further
complicates anomaly detection, leading to false positives or negatives.

Department of ISE, RNSIT 2024 – 2025 10

Hawk-Eye Smart Surveillance

2.11. Title: " AI-Driven Anomaly Detection for Cybersecurity: A Survey, Taxonomy,
and Open Issues"

Author: Vaishnavi Kho , Pratiksha Raut ,Neha Hake

Year:2023
Abstract: This paper presents a comprehensive survey on AI-driven anomaly detection
techniques in the context of cybersecurity. It categorizes existing approaches into
statistical, machine learning, and deep learning-based models. The paper also introduces
a taxonomy that classifies techniques based on their learning paradigm, data type, and
application domain. Furthermore, it discusses the strengths and limitations of these
approaches, outlines current challenges, and identifies future research directions to
enhance the effectiveness of anomaly detection systems in cybersecurity.
Limitations: This paper highlights several limitations in AI-driven anomaly detection
within cybersecurity. Chief among them is the scarcity of labeled datasets, which poses
a challenge for supervised learning models. Additionally, the dynamic and evolving
nature of cyber threats introduces concept drift, making static models less effective over
time. The study also notes the vulnerability of these systems to adversarial attacks,
which can deceive models into misclassifying threats. Furthermore, scalability remains
a concern, particularly when deploying AI models in large-scale, real-time
environments.

2.12. Title: " Smart Surveillance System for Anomaly Recognition"

Author: Praveen Kumar Thopalle.

Year:2024
Abstract: Proposes a smart surveillance system that autonomously detects anomalies
without human involvement using deep neural networks. It includes object detection,
tracking/discrimination, and anomaly recognition modules. The system is deployed as a
web app and sends alerts with snapshots via email.
Limitations: While the system proposed in this paper offers robust real-time detection
capabilities, it implicitly carries several limitations. The reliance on models like
YOLOv4, VGG-16, and LSTM means that the system is heavily dependent on the
accuracy and generalizability of these pretrained networks. The paper does not explicitly
address how the model performs in vastly different or unseen environments, suggesting

Department of ISE, RNSIT 2024 – 2025 11

Hawk-Eye Smart Surveillance

a potential limitation in adaptability. Furthermore, the model was trained primarily on

the UCF-Crime dataset, which might limit its effectiveness in detecting less common or
novel anomalies not present in the training data. There is also no discussion on handling
occlusions, poor lighting, or camera movement, which are common in real-world
surveillance scenarios.

2.13. Title: " A Comprehensive Review on Deep Learning-Based Methods for Video
Anomaly Detection "

Author: Umesh Chandra Pati, Rashmiranjan Nayak

Year:2024
Abstract: Introduces a smart surveillance system using YOLOv8 for real-time crowd
and threat detection. Aims to automate surveillance in public and critical infrastructure
environments. Demonstrates high accuracy and speed of YOLOv8 in object detection.
Limitations: This survey identifies several key limitations in the field. One of the most
pressing issues is the ambiguous and context-dependent definition of what constitutes
an anomaly, making model training and evaluation inherently difficult. The lack of
large-scale labeled datasets further restricts the development of supervised learning
approaches, pushing researchers toward unsupervised methods which may lack
precision. Additionally, human behavior is inherently complex and variable, which adds
to the difficulty in distinguishing normal from abnormal activities. The review also notes
that many current models do not scale well to real-time deployment, and that most
approaches struggle with interpretability, limiting their trustworthiness in sensitive
security applications.

2.14. Title: " Smart Surveillance Systems Using YOLOv8: A Scalable Approach for
Crowd and Threat Detection "

Author: Pallavi Chitte, Atharva Shanware, Pranit Jadhav.

Year:2024
Abstract: Introduces a smart surveillance system using YOLOv8 for real-time crowd
and threat detection. Aims to automate surveillance in public and critical infrastructure
environments. Demonstrates high accuracy and speed of YOLOv8 in object detection.
Limitations: Although the proposed system leverages the powerful YOLOv8 model for
real-time threat and crowd detection, it inherits some limitations from its underlying

Department of ISE, RNSIT 2024 – 2025 12

Hawk-Eye Smart Surveillance

architecture. The paper does not delve into how the model handles scenarios involving
low lighting, heavy occlusion, or complex human interactions, which are common in
surveillance footage. Moreover, its effectiveness is tied to the quality and diversity of
training data. If the training dataset lacks representation of certain threat types or
behaviors, the system may fail to detect them. The approach also assumes that threat
objects are always visually identifiable, overlooking non-visual cues that might be
critical in real-world security contexts.

2.15. Title: " Convolutional Transformer based Dual Discriminator Generative

Adversarial Networks for Video Anomaly Detection "

Author: Moshira Abdalla.

Year:2024
Abstract: Proposes an unsupervised framework (CT-D2GAN) for video anomaly
detection using a convolutional transformer and dual discriminators. Captures spatial
and temporal features for future frame prediction and uses prediction error for anomaly
detection.
Limitations: While CT-D2GAN presents an innovative unsupervised approach for
capturing spatiotemporal patterns in video data, it faces several limitations. The paper
points out that existing models often fail to effectively balance local (frame-level) and
global (temporal) consistency, and CT-D2GAN aims to address this gap. However,
despite improvements, the model’s complexity may hinder real-time performance,
especially on resource-constrained devices. Furthermore, unsupervised methods like
this often struggle with precision due to the lack of ground truth during training. There
is also no discussion on how the model adapts to different environments or types of
anomalies, suggesting potential limitations in generalizability and deployment.

2.16. Title: " Anomaly Detection and Prediction of Human Actions in a Video
Surveillance Environment "

Author: Anwaar Ulhaq, Muaz Al Radi.

Year:2024
Abstract: Proposes a hybrid model that combines deep learning-based feature
extraction with machine learning-based anomaly classification for human action

Department of ISE, RNSIT 2024 – 2025 13

Hawk-Eye Smart Surveillance

detection and prediction. The method uses spatiotemporal features and context
awareness for improved accuracy, achieving over 92% F1-score on benchmark datasets.
Limitations: While CT-D2GAN presents an innovative unsupervised approach for
capturing spatiotemporal patterns in video data, it faces several limitations. The paper
points out that existing models often fail to effectively balance local (frame-level) and
global (temporal) consistency, and CT-D2GAN aims to address this gap. However,
despite improvements, the model’s complexity may hinder real-time performance,
especially on resource-constrained devices. Furthermore, unsupervised methods like
this often struggle with precision due to the lack of ground truth during training. There
is also no discussion on how the model adapts to different environments or types of
anomalies, suggesting potential limitations in generalizability and deployment.

2.17. Title: " Video Anomaly Detection for Smart Surveillance: A Deep Learning
Approach "

Author: Bharathkumar Ramachandra.

Year:2024
Abstract: Introduces an unsupervised deep learning framework using autoencoders to
model normal behavior and detect anomalies based on reconstruction errors, eliminating
the need for labeled data.
Limitations: The proposed unsupervised autoencoder-based model demonstrates strong
potential, but is limited by several environmental factors. The reconstruction-based
approach is sensitive to disturbances such as camera shake, sudden illumination
changes, and occlusions, all of which can lead to high false positive rates. Although
dropout and skip connections are used to enhance performance, the system’s reliance on
frame-level appearance may not fully capture dynamic temporal anomalies.
Additionally, while the system operates at a viable 15 fps, it may not meet the demands
of real-time applications in high-throughput environments without further optimization.

2.18. Title: " A Survey on Anomalous Human Activity Recognition Using Video
Analytics "

Author: Moshira Abdalla, Michael J. Jones

Year:2024

Department of ISE, RNSIT 2024 – 2025 14

Hawk-Eye Smart Surveillance

Abstract: A comprehensive review of over 60 papers on video-based anomaly detection

using supervised, semi-supervised, and unsupervised techniques. Discusses evaluation
metrics, datasets, and emerging trends.
Limitations: This comprehensive survey paper identifies multiple limitations across
current methodologies. Chiefly, there is a consistent problem with class imbalance and
the limited availability of labeled anomaly data, which restricts the effectiveness of
supervised models. Unsupervised methods, while more flexible, often sacrifice
precision. The survey also highlights that many models are highly sensitive to changes
in scene composition, lighting, and camera angle, limiting their transferability.
Scalability remains an issue as most models are designed for controlled datasets and
struggle when applied to large, unstructured video archives. Interpretability and the
integration of multimodal data are also underdeveloped areas.

2.19. Title: " Deep Learning for Anomaly Detection in Surveillance Videos: Challenges
and Solutions "

Author: Abdelhafid Berroukham.

Year:2024
Abstract: Proposes a two-stage framework using ResNet-50 for feature extraction and
clustering for anomaly detection. Discusses challenges in interpretability, scalability,
and environmental variability.
Limitations: This paper discusses the application of a two-stage framework but also
details the inherent challenges of deploying deep learning models in surveillance. One
major limitation is the interpretability of deep networks, which is only partially
mitigated using visualization tools like Grad-CAM. Additionally, high computational
requirements are addressed through pruning and quantization, but these optimizations
may degrade model performance. The system's effectiveness is also influenced by
environmental conditions such as lighting and weather, and its adaptability to edge
devices like drones or embedded cameras is still a work in progress. These limitations
pose significant barriers to widespread deployment.

2.20. Title: " Predictive Analytics for Human Behavior in Surveillance Videos Using AI
Techniques "

Department of ISE, RNSIT 2024 – 2025 15

Hawk-Eye Smart Surveillance

Author: Khalid Housni

Year:2024
Abstract: Explores proactive behavior prediction using AI by combining rule-based
models, behavior modeling, scene understanding, and temporal analysis. Demonstrates
a 5-second average lead time for anomaly prediction.
Limitations: This paper discusses the application of a two-stage framework but also
details the inherent challenges of deploying deep learning models in surveillance. One
major limitation is the interpretability of deep networks, which is only partially
mitigated using visualization tools like Grad-CAM. Additionally, high computational
requirements are addressed through pruning and quantization, but these optimizations
may degrade model performance. The system's effectiveness is also influenced by
environmental conditions such as lighting and weather, and its adaptability to edge
devices like drones or embedded cameras is still a work in progress. These limitations
pose significant barriers to widespread deployment.

Department of ISE, RNSIT 2024 – 2025 16

 Chapter 3
PROBLEM STATEMENT

Traditional surveillance systems heavily rely on continuous human monitoring, which is time-
consuming, prone to error, and inefficient, especially when managing multiple video feeds.
Human operators may miss critical events due to fatigue, distraction, or information overload,
resulting in delayed responses to security threats. Moreover, conventional systems lack the
intelligence to automatically detect unusual behavior, unattended objects, or potential threats in
real time.

There is a growing need for an automated, intelligent surveillance solution that can analyze
video footage in real time, detect anomalies, and interpret human behavior without constant
human supervision. Such a system should be capable of identifying events like loitering,
aggressive movements, or abandoned objects and alerting security personnel instantly to
prevent or respond to incidents swiftly.

This project, Hawk-Eye: AI-Powered Smart Surveillance with Anomaly Detection &
Behavior Analysis, addresses this need by leveraging advanced AI and machine learning
technologies such as YOLOv8 for real-time object detection, MediaPipe Pose for behavior
tracking, and LSTM-based models for anomaly detection. The goal is to enhance situational
awareness, reduce reliance on manual monitoring, and improve overall security effectiveness
across various high-risk environments

Department of ISE, RNSIT 2024 – 2025 17

 Chapter 4
OBJECTIVES
• Develop an intelligent AI-based smart surveillance system that enhances traditional
video monitoring by integrating advanced computer vision and machine learning
techniques. The system aims to automate the detection of objects, analyze human
behavior, and identify abnormal or suspicious activities in real time, thereby improving
situational awareness and reducing reliance on manual surveillance efforts.
• Implement high-accuracy real-time object detection using YOLOv8, a state-of-the-
art deep learning model known for its speed and precision. The objective is to detect and
classify various objects commonly found in surveillance footage—such as people,
vehicles, bags, and other potentially suspicious items—enabling continuous tracking
and scene understanding.
• Analyze human behavior patterns through pose estimation using MediaPipe Pose,
which extracts skeletal keypoints from individuals in the video. This allows the system
to recognize and differentiate between different actions such as walking, standing,
falling, loitering, or displaying aggressive movements. The objective is to provide
behavior insights that can be used to flag unusual or potentially dangerous activities.
• Design and train an LSTM-based anomaly detection model capable of learning
temporal patterns of normal activities in surveillance video and identifying deviations
from these learned behaviors. The system will automatically flag anomalies like sudden
movements, crowd disturbances, or the presence of an unattended object, contributing
to proactive threat detection and response.
• Integrate a comprehensive web-based monitoring dashboard using Flask or
Streamlit, allowing security personnel to visualize live video streams or uploaded
footage, view flagged anomalies and behaviors, and receive instant alert notifications.
The goal is to create a user-friendly interface that enables efficient surveillance
management and quick reaction to potential security incidents.
• Ensure scalability and adaptability of the system so it can be deployed in a wide
range of environments such as public spaces, transportation hubs, office buildings, and

Department of ISE, RNSIT 2024 – 2025 18

Mini Project Report QuizMe : A MERN Based Educational Quiz
System
educational institutions. The system should be able to handle different camera inputs,
work with varying video qualities, and be extensible to support future upgrades like
multi-camera coordination and audio-based event detection.

Department of ISE, RNSIT 2024 – 2025 18

 Chapter 5
METHODOLOGY

SYSTEM IMPLEMENTATION
This chapter elaborates on the systematic implementation of the Smart Surveillance System. It
encompasses the structural design, the integration of key modules, and the deployment strategies
employed for anomaly detection, behavioral analysis, and unattended object monitoring using
deep learning and computer vision techniques.

5.1 Structural Design

Figure 5.1: Structural Chart of the proposed system

Department of ISE, RNSIT 2024 – 2025 19

Hawk-Eye Smart Surveillance

Figure 5.1: Structural Chart of the Proposed System

The structural chart of the proposed system outlines the sequential flow of operations and the
interaction between various components of the Smart Surveillance System. The process begins
with data input and ends with the classification and alert generation based on detected anomalies
or behaviors.

Step-wise Explanation:

1. Video Data Acquisition:

o The process starts with uploading pre-recorded surveillance footage into the
system via the Streamlit interface. The video acts as the primary input and can be
from any standard CCTV feed in MP4 or AVI format.

o The system supports batch upload and segment-wise processing for enhanced
performance and better memory management.

2. Preprocessing and Frame Extraction:

o The uploaded video is broken down into individual frames using OpenCV. This
allows frame-by-frame analysis which is essential for temporal modeling and
behavior tracking.

o Frames are resized and normalized to maintain uniformity and ensure compatibility
with the trained model dimensions.

o Background subtraction and noise removal techniques are applied to reduce

irrelevant data and enhance focus on foreground objects.

3. Feature Extraction:

o For each frame, spatial features (object boundaries, contours, and textures) are
extracted using convolutional layers.

o Temporal consistency is captured using frame differencing and motion history

images to track the progression of events in the video sequence.

4. Anomaly Detection (Deep Learning-based):

Department of ISE, RNSIT 2024 - 2025 20

Hawk-Eye Smart Surveillance

o The system utilizes a Convolutional Autoencoder trained on normal behavior

patterns to identify anomalies.

o During inference, each frame is passed through the encoder-decoder architecture.

The reconstruction error is computed between the original and reconstructed frame.

o Frames with high reconstruction error beyond a set threshold are flagged as
potential anomalies, indicating a deviation from learned normal behavior.

5. Behavior Analysis:

o The system evaluates behavior by analyzing frame sequences over time. It clusters
patterns based on movement and pose dynamics without explicitly labeling
specific activities.

o The temporal analysis is supported using optical flow and recurrent observations
of keypoints and frame changes.

o This module helps the system understand whether the observed action is consistent
with regular movement patterns or indicates an unusual behavioral pattern.

6. Unattended Object Detection:

o A tracking mechanism is implemented to identify stationary objects across

multiple consecutive frames.

o If an object remains in the same position beyond a predefined time threshold, the
system logs it as an unattended object, potentially indicating a security risk.

7. Alert Generation and Output Interface:

o Alerts and flagged anomalies are displayed on the Streamlit-based dashboard.

o Users can view a summary of anomalies detected, review timestamps of flagged

frames, and download reports.

o The interface also includes visualization tools such as reconstruction error graphs,
frame differencing heatmaps, and behavior timelines.

Department of ISE, RNSIT 2024 - 2025 21

Hawk-Eye Smart Surveillance

5.2 Video Frame Analysis Pipeline

The pipeline is a crucial backbone of the system, designed to ensure smooth operation from data
ingestion to final output generation. It includes the following stages:

1. Frame Loader:

• Splits video into frames using OpenCV’s VideoCapture class.

• Frames are stored in a queue for asynchronous processing to avoid memory overflow.

2. Preprocessing Unit:

• Performs:

o Resizing to 64x64 or 128x128 resolution

o Grayscale conversion

o Noise reduction using Gaussian Blur

• Converts each frame into a format suitable for model input (e.g., numpy arrays or tensors).

3. Autoencoder-Based Anomaly Detector:

• A deep learning model trained using normal surveillance footage.

• Utilizes Mean Squared Error (MSE) between input and reconstructed output to determine
anomaly levels.

• Frames above a defined error threshold are classified as anomalous.

4. Object Tracker (for unattended object detection):

• Applies object detection (e.g., background subtraction + contour tracking) to locate static
items.

• Uses a timer to check how long each object has remained in place without moving.

• If the duration exceeds the safety limit, an alert is triggered.

5. Behavior Analysis Module:

• Leverages spatial-temporal features across frames.

Department of ISE, RNSIT 2024 - 2025 22

Hawk-Eye Smart Surveillance

• Implements motion analysis to interpret patterns over time.

Behavioral heatmaps and trajectory maps are created to visualize movement consistency or
irregularity.

5.3 Deep Learning Architecture: Convolutional Autoencoder

The core component for anomaly detection is a Convolutional Autoencoder, trained to

reconstruct normal surveillance video frames. The key design aspects are:

Encoder:

• Series of convolutional and pooling layers to compress input frame into a latent
representation.

• Captures spatial semantics and structural regularities of the scene.

Decoder:

• Transposed convolutional layers to reconstruct the input frame from the encoded latent
vector.

• Trained to minimize reconstruction loss using MSE.

Training Strategy:

• Dataset: Avenue or UCSD dataset used for training.

• Loss Function: Mean Squared Error

• Optimizer: Adam optimizer with learning rate scheduling

• Epochs: Trained over 50–100 epochs for stable convergence.

5.4 User Interface Design (Streamlit-based)

To enhance usability, a web interface built using Streamlit is provided, which includes:

• Upload section for selecting surveillance videos

• Progress bar and status indicators for frame processing

Department of ISE, RNSIT 2024 - 2025 23

Hawk-Eye Smart Surveillance

• Live graph showing reconstruction error across frames

• Playback of anomalous frames with labels

• Downloadable report of all flagged events

Department of ISE, RNSIT 2024 - 2025 24

Hawk-Eye Smart Surveillance

Chapter 6

SYSTEM DESIGN
6.1 System Environment

The system architecture for the MERN-based Quiz Web Application is designed to ensure a
seamless interaction between the front-end and back-end, providing a robust, scalable, and
efficient platform. Here’s an overview of the architecture:

The system environment of a Smart Surveillance and Anomaly Detection System is the
physical and software infrastructure that ensures efficient, secure, and reliable operation.

• Hardware:
Powerful enough to handle video frame extraction, deep learning-based anomaly
detection, behavior analysis, and object monitoring efficiently. A system with a multi-core
processor, sufficient RAM, and preferably a dedicated GPU for faster model inference is
recommended.

• Software:
Compatible with machine learning, computer vision, and video processing libraries (like
TensorFlow, OpenCV, and Streamlit). Provides a secure and stable environment for
sensitive video data handling and anomaly detection processes.

• Network:
Reliable and with sufficient bandwidth for uploading analysis results, generating alerts,
and optionally sending notifications. If remote monitoring is enabled, it should support
smooth data transfer without interruptions.

6.2 System Components

• Anomaly Detection Algorithm:
The core component that detects unusual patterns or abnormal activities by analyzing
video frames using a convolutional autoencoder model.

Department of ISE, RNSIT 2024 - 2025 25

Hawk-Eye Smart Surveillance

• Behavior Analysis Module:

Classifies human actions in video frames (e.g. walking, standing, running) based on
extracted spatial-temporal features.

• Unattended Object Monitoring:

Detects and tracks objects left in one place for too long and raises alerts when thresholds
are exceeded.

• Data Collection and Storage:

Collects and stores processed video frames, anomaly logs, and behavior classification
results securely in local or cloud storage.

• User Interface:
Provides an interactive, user-friendly dashboard (built with Streamlit) to view live or
processed results, alerts, and activity logs.

• Administration Interface:
Enables administrators to manage video datasets, configure detection thresholds, and
monitor system health and performance securely.

• Security:
Protects video data and system configurations through encryption, access controls, and
secure communication channels.

• Performance:
Ensures efficient frame extraction, model inference, and result generation to process
surveillance data quickly and reliably.

• Scalability:
Supports the integration of larger video datasets, additional camera feeds, and expanded
detection functionalities as required.

• Maintainability:
Designed to be easily updated with new models, detection parameters, and UI
enhancements as technology and security needs evolve.

Department of ISE, RNSIT 2024 - 2025 26

Hawk-Eye Smart Surveillance

6.3 System Interfaces

• The system should have user-friendly, intuitive, and secure interfaces.

• The user interface must be easy to use and navigate for uploading videos and viewing
analysis results.

• The administration interface must be secure and allow administrators to manage video
data, configure detection settings, and monitor system performance.

6.4 System Constraints

• The system must be powerful, secure, accurate, scalable, maintainable, and efficient.

• The system must be able to collect, process, and store large volumes of video data.

• The system must be able to operate in various physical and network environments.

6.5 System Assumptions

• The anomaly detection and behavior analysis algorithms are accurate and reliable.

• The data collection and storage system is secure and ensures privacy protection.

• The user interface is user-friendly, stable, and reliable.

• The administration interface is secure and allows effective system management.

• The system will remain scalable and maintainable for future upgrades and
enhancements.

6.6 Data Flow Diagram

1. Video Input:
o User uploads pre-recorded video files via UI.
2. Preprocessing:
o Video is split into frames.
o Frames are resized, normalized, and prepared for analysis.
3. Anomaly Detection:
o Frames passed through CAE.

Department of ISE, RNSIT 2024 - 2025 27

Hawk-Eye Smart Surveillance

o Reconstruction error calculated.

o Anomalies flagged if error > threshold.
4. Behavior Analysis:
o Actions detected by classifying frame sequences.
o Action labels stored with timestamps.
5. Unattended Object Monitoring:
o Static objects tracked frame by frame.
o Timer triggered if object remains static too long.
o Flagged as unattended if timer exceeds set duration.
6. Result Generation:
o All analysis results compiled into reports.
o Displayed on UI and optionally sent as notifications.

Figure 6.1: Flow Chart of The System

Department of ISE, RNSIT 2024 - 2025 28

Hawk-Eye Smart Surveillance

Chapter 7

CONCLUSION
7.1 Conclusion

Hawk-Eye: AI-Powered Smart Surveillance with Anomaly Detection & Behavior Analysis
redefines the way surveillance systems function by integrating advanced artificial intelligence
techniques to detect, interpret, and respond to unusual activities. Traditional surveillance heavily
relies on manual monitoring, which is often inefficient and error-prone, especially in
environments with multiple video feeds. Hawk-Eye overcomes these limitations by automating
video analysis using machine learning models that are capable of detecting frame-level anomalies,
tracking unattended objects, and analyzing human behavior without requiring manually labeled
data. This makes the system highly scalable and adaptable for real-world security applications.

At the heart of Hawk-Eye lies a deep learning-based ConvLSTM autoencoder model, which learns
typical patterns of motion and activity from training video footage. Any significant deviation from
these patterns is flagged as anomalous, allowing the system to detect events such as sudden
movements, unauthorized presence, or unexpected object placement. This unsupervised learning
approach not only reduces the dependency on manually annotated datasets but also enables the
system to adapt to different environments with minimal retraining. Additionally, the unattended
object detection module ensures that any object left in a fixed location beyond a predefined
duration is identified and flagged, enhancing threat detection in sensitive areas such as airports,
offices, and public venues.

Although Hawk-Eye currently processes pre-recorded video files rather than real-time feeds, it
sets a strong foundation for future enhancements. Planned upgrades such as real-time video stream
support, multi-camera integration, and audio-based anomaly detection will further elevate its
capabilities. By transforming passive video monitoring into an intelligent, automated security tool,
Hawk-Eye demonstrates how AI can significantly improve situational awareness, response times,
and overall safety in both public and private spaces.

Department of ISE, RNSIT 2024 - 2025 29

Hawk-Eye Smart Surveillance

REFERENCES

[1]. Chris, E., Olabiyi, W., Favor, B., & Ramon, T. (2023). "AI-powered surveillance systems
and anomaly detection." Available at: ResearchGate
[2]. (2023). "Deep Learning-Based Anomaly Detection in Video Surveillance." Available at:
MDPI
[3]. (2023). "AI-Powered Anomaly Detection Techniques." Available at: ResearchGate
[4]. (2023). "Revolutionizing Video Surveillance: AI-Powered Anomaly Detection."
Available at: ResearchGate
[5]. (2023). "AI-powered threat detection in surveillance systems: A real-time data processing
framework." Available at: ResearchGate
[6]. (2024). "Towards a Conceptual Framework for AI-Driven Anomaly Detection in Smart
City IoT Networks." Available at: ScienceDirect
[7]. (2023). "AI-Based Anomaly Detection for Real-Time Cybersecurity." Available at:
ResearchGate
[8]. (2023). "Deep Learning-Based Anomaly Detection in Video Surveillance: A Survey."
Available at: ResearchGate
[9]. (2023). "AI-Powered Anomaly Detection for Real-Time Performance Monitoring in
Cloud Systems." Available at: ResearchGate
[10]. (2023). "Video Anomaly Detection System Using Deep Convolutional and
Recurrent Neural Networks." Available at: ScienceDirect
[11]. (2023). "AI-Driven Anomaly Detection." Available at: ResearchGate
[12]. (2023). "Smart Surveillance System for Anomaly Recognition." Available at:
ResearchGate
[13]. (2023). "A Comprehensive Review on Deep Learning-Based Methods for Video
Anomaly Detection." Available at: ScienceDirect
[14]. (2023). "Smart Surveillance Systems Using YOLOv8: A Scalable Approach for
Crowd and Threat Detection." Available at: ResearchGate
[15]. (2021). "Convolutional Transformer-Based Dual Discriminator Generative
Adversarial Networks for Video Anomaly Detection." Available at: arXiv

Department of ISE, RNSIT 2024 - 2025 30

Hawk-Eye Smart Surveillance

[16]. (2020). "Multiple Instance-Based Video Anomaly Detection Using Deep

Temporal Encoding Decoding." Available at: arXiv
[17]. (2020). "Continual Learning for Anomaly Detection in Surveillance Videos."
Available at: arXiv
[18]. (2020). "Video Anomaly Detection Using Pre-Trained Deep Convolutional Neural
Nets and Context Mining." Available at: arXiv
[19]. (2020). "Video Anomaly Detection Using Deep Convolutional Networks."
Available at: MDPI
[20]. (2020). "Anomaly Detection in Video Surveillance Using Deep Learning."
Available at: ScienceDirect

Department of ISE, RNSIT 2024 - 2025 31