Computer Vision
Shoplifting Detection System
A computer-vision pipeline for shoplifting detection that combines person detection, tracking, pose extraction, and live risk scoring.
YOLOv8DeepSORTMediaPipeFastAPI
Overview
Built from scratch around YOLOv8 for detection, DeepSORT for tracking, MediaPipe for pose, and a custom behavior engine.
Supports RTSP streams, webcams, and video files, with annotated output and automatic alert snapshots.
The real challenge was reducing false alerts in retail-like behavior rather than simply wiring models together.
Results
End-to-end pipeline from RTSP / webcam / video input to annotated alert output
Custom behaviour-risk scoring engine reducing false positives vs naive detection
Automatic alert snapshot capture on high-risk events
Key Outcomes
Generated annotated video with pose overlays, risk labels, and saved alert snapshots.
Focused heavily on behavior thresholds to separate normal shelf interaction from suspicious motion.
Outlined next steps toward better retail fine-tuning and transformer-based behavior modeling.
Need the full implementation story?
Walkthrough, private repo access, or implementation details available.