Real Time Theft Detection using AI

In this project, we present an advanced, real-time AI-based theft detection system designed to enhance surveillance in retail environments. The system architecture is engineered for high performance, moving beyond simple motion tracking by implementing a sophisticated, multithreaded pipeline. This pipeline combines deep learning for content analysis, computer vision for motion analysis, and an ensemble of machine learning models to accurately identify and flag suspicious activities. The real-time capability is achieved using a dedicated detection worker thread that processes frames asynchronously from a queue, ensuring that the main camera feed remains smooth and uninterrupted. The core of the system is a novel hybrid feature engineering approach. A pretrained ResNet-18 Convolutional Neural Network (CNN) is utilized through PyTorch, where its final classification layer is removed to transform it into a powerful deep feature extractor. This allows each video frame to be converted into a rich 512-dimensional content vector. At the same time, a Motion Analyzer module leverages OpenCV to perform parallel motion analysis. This module extracts eight distinct motion features by combining two techniques: background subtraction for detecting foreground objects and frame differencing for identifying pixel-level changes. These features include motion density, average pixel variation, and the area of the largest moving contour. The extracted 512 content features and 8 motion features are combined to form a single 520-dimensional feature vector. This vector is normalized using a standard scaling technique and then used to train an Ensemble Voting Classifier. The ensemble integrates four powerful machine learning models: Random Forest, Gradient Boosting, Multi-layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). The classifier operates using soft voting, where the probability outputs from all models are averaged to produce a more accurate and reliable prediction. To address the issue of class imbalance, where theft instances are significantly fewer than normal instances, the SMOTE technique is applied during preprocessing. This ensures that the dataset is balanced, improving the model’s ability to detect rare events effectively. Additionally, the alert system is carefully designed to minimize false positives. An alert is triggered only when multiple consecutive frames indicate suspicious activity and when the average confidence level exceeds a predefined threshold. A cooldown mechanism is also implemented to prevent repeated alerts within a short time frame. Once a suspicious event is confirmed, an instant notification is sent to security personnel via Telegram. The system also includes a Flask-based web dashboard that provides a real-time monitoring interface. This dashboard offers API endpoints to access detection statistics, view recent logs, and manage trained models. All configurations are centrally managed using a configuration file, ensuring flexibility and scalability. Overall, this multi-layered system delivers a robust, accurate, and scalable solution for automated theft detection in real-world environments. KEYWORDS : Real-Time Theft Detection, Computer Vision, Deep Learning, ResNet-18, Motion Analysis, Ensemble Learning, XGBoost, SMOTE, Video Surveillance, Flask Dashboard, Multithreading, Suspicious Activity Detection