Theft Guard AI: Real-Time Theft Detection System

About the Project

In the retail sector, shrinkage (theft) is a multi-billion dollar global problem for businesses. While physical security measures, alarm sensors, and cameras act as deterrents, they are usually insufficient to stop theft incidents the moment they occur. Having security personnel continuously monitor hundreds of hours of CCTV footage simultaneously is a massive operational cost and is highly prone to human error.

Theft Guard AI is an intelligent early-warning system that fully automates this process using modern artificial intelligence and Computer Vision algorithms. It connects directly to the existing IP cameras already installed in the store. By processing the live video feed with millisecond latency, it analyzes the suspicious behaviors, body language, and interactions of individuals inside the store.

Unlike traditional motion sensors, it understands not just the presence of motion, but the intent behind it. By detecting critical events—such as quickly hiding a product inside a pocket or bag—in real time, it instantly sends alerts to mobile devices.

Comprehensive Features and Modules

• Real-Time Skeleton and Behavior Analysis: Using YOLOv8 Pose Estimation and custom-trained classification models, the system extracts joint points on the human body. It instantly detects risky actions like “hand-to-pocket” concealment, suspicious bending over, or unauthorized entry into restricted areas with an accuracy rate of over 95%.
• Advanced Face Recognition and Watchlist: Thanks to the integrated Face Recognition algorithm, the faces of suspects who have previously shoplifted or caused trouble can be added to a watchlist. These individuals are detected the very first second they walk through the store doors, generating a high-priority red alert.
• Smart Geofencing (Virtual Boundaries): Virtual boundaries can be drawn around specific areas of the store (e.g., the expensive electronics aisle, behind the cash register, or staff room entrances). The system’s sensitivity level is automatically maximized at the slightest suspicious action in these areas.
• Instant and Multi-Channel Notification System: When a theft is detected, the suspect’s cropped photo, risk score, and location information are automatically dropped into the store manager’s or security guard’s Telegram app within seconds. Detailed email reports (PDF) can also be generated for management.
• DVR and Evidence Management: Every analyzed suspicious event, along with video clips, timestamps, camera names, and security scores, is stored end-to-end encrypted in a local SQLite or PostgreSQL database. Searching historical events takes only seconds.
• Live Monitoring and Analytics Dashboard: It features a modern, dark-mode supported web interface built with Next.js. It offers the ability to watch the live feed of all cameras, extract weekly or monthly theft attempt statistics, and manage camera statuses.

Why and How is it Different from Competitors?

Theft Guard AI is designed entirely with Edge Computing logic, meaning it runs directly on local servers or computers (on-premise) at the store. This provides two major advantages:

1. Zero Cloud Costs: Video feeds are not sent over the internet to expensive cloud servers for analysis. It consumes no internet bandwidth and eliminates the cost of transferring gigabytes of video data.
2. Flawless Data Privacy: Customer images never leave the store. The system features face blurring or anonymization capabilities to be fully compliant with personal data protection laws like GDPR.

Other commercial software on the market is usually sold alongside proprietary hardware, forcing businesses to buy specific brands of cameras (vendor lock-in). Theft Guard AI can integrate with any brand and model of standard camera that supports the RTSP protocol.

Who is it Ideal For?

• Retail Store Chains and Supermarkets: Those wanting to minimize shelf losses in cosmetics, clothing, and food retail.
• Jewelers and Electronics Stores: Specific businesses selling high-value, small-sized items that require maximum security.
• Shopping Malls and Logistics Warehouses: Security teams aiming to monitor suspicious wandering in large areas.
• SMBs: Any business wanting to make their existing cameras “smart” with an open-source AI instead of purchasing expensive proprietary hardware and software licenses.

Architecture and Technical Infrastructure

The entire system is built asynchronously, conforming to a microservices architecture.

• AI & Backend Engine: Developed using Python and FastAPI. YOLOv8 and OpenCV libraries are accelerated with TensorRT, running on the GPU without FPS loss. Communication with cameras is isolated using multiprocessing structures.
• Real-Time Communication: Events and video frames are transmitted to the Frontend via low-latency WebSockets.
• Frontend / Management Panel: A highly performant and responsive interface is created using Next.js 14, React, TailwindCSS, and Zustand.
• Environment Support: Can easily be deployed on Windows, Linux, or Docker environments.

GitHub Repository: github.com/vahapogut/Theft-Detection