Retail Theft Detection
Real-time computer vision system for loss prevention
Researcher
Detection Pipeline
Background
Retail theft costs businesses billions annually. Most loss prevention relies on reactive CCTV review after an incident. This research explores a proactive, automated approach using deep learning for real-time detection and tracking.
Technical Approach
YOLOv8 was chosen for its balance of speed and accuracy — it runs at real-time fps on consumer-grade hardware. DeepSORT extends single-frame detection to full trajectory tracking by combining Kalman filter predictions with deep appearance features (Re-ID). The behavior analysis layer uses heuristic rules derived from LP domain knowledge: dwell time in high-theft zones, occlusion events near shelves, and unusual trajectory patterns.
Results
The system achieved high detection accuracy on the evaluation dataset. False positive rate was kept low through a confidence threshold tuned to minimize unnecessary alerts. The paper was accepted and published in the IJSREM peer-reviewed journal — my first academic publication.
Learnings
This project taught me the gap between academic benchmarks and real-world deployment. Lighting variation, camera angle, and occlusion are far harder in-the-wild than in curated datasets. The next iteration would incorporate domain adaptation and a confidence calibration layer.