Real-time Theft Detection through

PROJECT Intelligent Video Analysis

TITLE
Phase 1
Group No : 6
GUIDE: Manjusha R.
Review 1

S.No Name Roll Number

1 Aditya Krishnan Mohan CB.EN.U4CSE21202
2 Parthiv Chandran CB.EN.U4CSE21245
3 Anish Nandhan E CB.EN.U4CSE21204
4 Sowrabh S.kumar CB.EN.U4CSE21058

Date: 9/24/2024 Team No: 6

 Guide’s Approval Mail

Date: 9/24/2024
 Introduction
● Retail Losses: Shoplifting is a major concern for retailers worldwide, causing
significant financial losses every year.
● Current Surveillance Methods: Traditional shoplifting prevention relies on manual
monitoring of CCTV footage, which is both labor-intensive and prone to human
error.
● Advancements in AI: Recent advancements in artificial intelligence (AI) and
computer vision provide the opportunity to automate the detection of shoplifting
activities.
● Objective of the Project: This project focuses on building a real-time shoplifting
detection system using video analysis, employing cutting-edge techniques such as
object detection, tracking, and behavior analysis to identify suspicious activities.

Date: 9/24/2024
 Motivation
● High Cost to Retailers: Global retailers face billions of dollars in losses
annually due to shoplifting, making it a critical issue for the industry.
● Limitations of Human Surveillance: Manual surveillance systems rely
heavily on human operators, who may miss incidents due to fatigue,
distraction, or workload.
● Need for Real-Time Solutions: Detecting potential theft as it happens
is crucial for preventing losses. Automated systems can identify
suspicious activities in real-time and alert security personnel instantly.
● Efficiency and Accuracy with AI: AI-based video analysis systems can
continuously monitor large areas, providing higher accuracy, faster
response times, and reducing the need for constant human supervision.
This helps in preventing theft before it occurs, rather than just analyzing
it after the fact.
● Scalable and Adaptable: An AI-powered solution is scalable to different
store sizes and environments, making it adaptable to a wide range of
retail settings.

Date: 9/24/2024
 Literature Survey
Study Title Author Year of publication Technique used Pros Cons

Simple tracking of
Luminance contrast, events, low-level
Fuentes, L.M., overlapping bounding Dependency on proper
Tracking-based event detection for CCTV syste 26 November 2004 boxes, blob trajectory
description of events
ms Velastin, S.A. lighting conditions
analysis based on blob positions
and density

Jerry Gao, Jingwen YOLOv5 for arson,

Shi ,Priyanka YOLOv7 for
Utilizes multiple YOLO
Balla ,Akshata burglary, YOLOv6 Reliance on high-
Camera-Based Crime Behavior Detection and C 19 May 2024 models, integrates a
lassification Sheshgiri ,Bocheng for vandalism, resolution videos
Twilio-based alert system
Zhang ,Hailong Yu OpenCV decision
and Yunyun Yang algorithm

Tasriva Sikandar, Effective facial

Struggles with detecting
Kamarul Hawari CNNs (VGGNet, recognition even in
nuanced behaviors,
Ghazali & 11 December 2018 ResNet), GMM for challenging conditions,
ATM Crime Detection Using Image Processing unable to adapt to
Mohammad Fazle motion detection reduced false positives in
evolving criminal tactics
Rabbi complex environments

Date: 9/24/2024
 Literature Survey
Study Title Author Year of publication Technique used Pros Cons

Effective for detecting

3D Convolutional criminal behaviors over Limited dataset diversity,
A Study on Abnormal Behavior Detection in CCTV Yangsun Lee June 2023 Neural Networks multiple frames, accurate scalability challenges in
Images through the Supervised Learning Model of (C3D) classification of abnormal
Deep Learning large-scale systems
activities

Kishan Bhushan Real-time detection with Struggles with real-time

Sahay , Bhuvaneswari Spatio-Temporal (ST) spatio-temporal data
Balachander, B. and Deep detection in cluttered
A real time crime scene intelligent video surveil Jagadeesh , G. Anand 1 September 2022 Reinforcement Neural
analysis, efficient environments, scalability
lance systems in violence detection framework processing of violent challenges
using deep learning techniques Kumar , Ravi Kumar, Networks (DRNN)
L. Rama Parvathy actions

YOLO algorithm, High accuracy in Struggles with detecting

K. Ganagavalli & V. LSTM-CNN hybrid detecting vandalism, subtle behaviors,
YOLO-Based Anomaly Activity Detection for 23 April 2024
Human Behavior Analysis Santhi model for feature integrates LSTM for performance degrades in
extraction action classification poor lighting

Date: 9/24/2024
 Literature Survey
Study Title Author Year of publication Technique used Pros Cons

VSD for video Gap between low level

Zheng Xu,
3 December parsing and More effective organization video features and high-
Video structured description technology based i Chuanping Hu, and retrieval of video data level semantic
ntelligence analysis of surveillance videos for p Lin Mei 2015 AlexNet CNN for
ublic security applications identification understanding

Kishan Bhushan
Sahay , Bhuvaneswari Struggles with detecting
YOLOv5, Faster R- Real-time performance,
Balachander, B. small or occluded
Image and video-based crime prediction using o Jagadeesh , G. Anand 1 September 2022 CNN, SSD for object high detection speed with
objects, balancing
bject detection and deep learning detection YOLOv5
Kumar , Ravi Kumar, accuracy with speed
L. Rama Parvathy

Date: 9/24/2024
 Research Gap
● Dependency on Proper Lighting and Environmental Conditions: Many systems, especially those
using luminance contrast-based segmentation or YOLO-based detection, are highly dependent on good
lighting and ideal environmental conditions, limiting their effectiveness in real-world low-light or cluttered
environments.
● Scalability Issues: Deep learning models, such as C3D and YOLO, face challenges when applied to
large-scale surveillance networks with multiple cameras. These systems often struggle to maintain real-
time detection efficiency across extensive surveillance setups.
● Limited Dataset Diversity and Generalization: Current models are trained on datasets that lack
diversity in crime scenarios, environmental factors, and scene types. This limits the models' ability to
generalize effectively across different real-world settings, resulting in reduced detection accuracy.
● Difficulty Handling Complex or Subtle Human Behaviors: Systems often fail to detect nuanced or
ambiguous behaviors, such as gradual or subtle actions, which can be precursors to criminal activity.
This reduces their effectiveness in identifying behaviors that do not fit clear-cut patterns.
● Inability to Adapt to Evolving Criminal Tactics: Existing systems lack the flexibility to adapt to new or
evolving criminal behaviors, such as the use of disguises or coordinated actions. This rigidity can cause
these models to become outdated and ineffective over time.

Date: 9/24/2024
 Problem Statement
The goal is to design a real-time theft detection system that can:

● Balance between accuracy and speed for real-time surveillance.

● Handle evolving criminal behavior and nuanced actions across different
environments.

By addressing these gaps, the proposed solution aims to enhance security in

public spaces by reducing response time and improving detection accuracy
across complex, large-scale surveillance systems.

Date: 9/24/2024
 Module Description
● Video Input Module: Captures live video feeds from surveillance cameras in real-time. Handles the ingestion of
video streams from various camera sources in the retail environment, ensuring compatibility with different
formats and resolutions.
● Preprocessing Module: Enhances video quality and extracts frames for further analysis. Applies techniques like
noise reduction, frame extraction, and video trimming to optimize data for accurate detection, handling low-light
environments and camera angles.
● Object Detection and Tracking Module: Detects and tracks objects, specifically individuals and items within the
video. Utilizes advanced object detection algorithms (e.g., YOLO, Faster R-CNN) to identify individuals, track
movements, and monitor interactions with merchandise.
● Behavior Analysis and Anomaly Detection Module: Analyzes detected behaviors for suspicious or anomalous
activities related to shoplifting. Flags potential incidents based on movement patterns, item handling, and
behavior characteristics using models of normal vs. suspicious behavior.
● Prediction and Alert Module: Predicts potential shoplifting incidents and sends alerts. Uses machine learning
models to analyze behavior patterns and sends real-time alerts to store personnel upon detecting an anomaly.
● Logging and Reporting Module: Maintains a record of flagged incidents and generates reports for store
management. Logs detected events, stores video footage, and generates reports for post-event analysis,
integrating with store systems for further action.
Date: 9/24/2024
 Role of each Student
1. Sowrabh will handle data collection and preprocessing, including gathering crime-related videos,
trimming, generating frames, resizing, and applying augmentations to prepare the dataset for
training.

2. Aditya will focus on developing and training object detection models like YOLOv5, YOLOv7, and
YOLOv6 for detecting crimes such as arson, burglary, and vandalism, ensuring accuracy and
efficiency.

3. Parthiv will be responsible for deploying the models and optimizing them for real-time detection,
integrating with APIs like Twilio and Gradio.

4. Anish will work on improving model performance through hyperparameter tuning and ensemble
techniques, ensuring better crime detection outcomes by integrating multiple models.

Date: 9/24/2024
 References
[1] J. Gao, J. Shi, P. Balla, A. Sheshgiri, B. Zhang, H. Yu, and Y. Yang, "Camera-Based Crime Behavior Detection and Classification," *Smart Cities*, vol. 7, no. 3, pp.
1169-1198, 2024. [Online]. Available: https://doi.org/10.3390/smartcities7030050

[2] L. M. Fuentes and S. A. Velastin, "Tracking-based event detection for CCTV systems," *Pattern Anal Applic*, vol. 7, pp. 356-364, 2004. [Online]. Available:
https://doi.org/10.1007/s10044-004-0236-z

[3] T. Sikandar, K. H. Ghazali, and M. F. Rabbi, "ATM crime detection using image processing integrated video surveillance: a systematic review," *Multimedia Systems*,
vol. 25, pp. 229-251, 2019. [Online]. Available: https://doi.org/10.1007/s00530-018-0599-4

[4] Y. Lee, "A Study on Abnormal Behavior Detection in CCTV Images through the Supervised Learning Model of Deep Learning," *Journal of Logistics, Informatics and
Service Science*, vol. 9, no. 2, pp. 196-209, 2022. [Online]. Available: https://doi.org/10.33168/LISS.2022.0212

[5] K. B. Sahay, B. Balachander, B. Jagadeesh, G. A. Kumar, R. Kumar, and L. R. Parvathy, "A real time crime scene intelligent video surveillance systems in violence
detection framework using deep learning techniques," *Computers and Electrical Engineering*, vol. 103, 2022. [Online]. Available:
https://doi.org/10.1016/j.compeleceng.2022.108319

[6] M. Boukabous and M. Azizi, "Image and video-based crime prediction using object detection and deep learning," *Bulletin of Electrical Engineering and Informatics*,
vol. 12, no. 3, pp. 1630-1638, 2023. [Online]. Available: https://doi.org/10.11591/eei.v12i3.S157

[7] Z. Xu, C. Hu, and L. Mei, "Video structured description technology based intelligence analysis of surveillance videos for public security applications," *Multimedia
Tools Appl*, vol. 75, pp. 12155-12172, 2016. [Online]. Available: https://doi.org/10.1007/s11042-015-3112-5

[8] K. Ganagavalli and V. Santhi, "YOLO-based anomaly activity detection system for human behavior analysis and crime mitigation," *SIViP*, vol. 18, Suppl. 1, pp. 417-
427, 2024. [Online]. Available: https://doi.org/10.1007/s11760-024-03164-7

Date: 9/24/2024
 THANK YOU

Date: 9/24/2024