Shoplifting Detection System

Deep learning-based video classification system for detecting shoplifting behavior in surveillance footage using 3D Convolutional Neural Networks (R3D-18). Analyzes 16-frame video clips at 224x224 resolution for binary classification of shoplifting vs normal activity. Trained on 183 videos (90 normal + 93 shoplifting).

The system processes surveillance video in real-time, extracting temporal features across consecutive frames to understand human actions and identify suspicious shoplifting behavior. The R3D-18 architecture, pretrained on the large-scale Kinetics-400 dataset, provides robust spatiotemporal feature extraction that generalizes well to retail surveillance scenarios.

Key Features

• R3D-18 (3D ResNet-18) pretrained on Kinetics-400
• 16-frame video clip analysis at 224x224
• Binary classification (shoplifting vs non-shoplifting)
• 183-video training dataset (80/20 split)
• CUDA GPU acceleration support
• Annotated output video with predictions and confidence scores