CHURI: Contour-Gated, Resource-Intelligent Retail Tracking System

Riya Waghmare; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 1574-1581

Most retail tracking solutions assume always-on cloud analytics, making basic counting a subscription-based service. For small stores (47.3% of establishments), 40-120/month per camera can exceed 2-6% profit margins, so tracking is disabled and CCTV stays passive. Yet retail is harder than benchmarks: overhead fisheye views need specialized line-crossing or top-down models, and accurate tracking often requires GPU-class inference (cloud or 8,000+ NVRs)-well beyond 200-500 small-store budgets despite occupancy needs for compliance and theft prevention. At this operating point, existing edge methods fail: standalone YOLO on Raspberry Pi 4B achieves 52.1% recall while saturating CPU at 82% for single-camera processing; classical BGS drifts without detector confirmation (MOG2: 40.5%, KNN: 54.8%); and sparse detection (YOLO every N frames) causes 35-48% ID-switch rates due to fragmented trajectories. We introduce CHURI, a detection-gated architecture for resource-constrained retail. CHURI employs a three-stage cascade: (i) lightweight monitoring via zone-BGS at 2 FPS to detect motion, (ii) gated verification via YOLO with an 8% invocation rate to confirm persons, and (iii) persistent tracking via contour-velocity propagation to sustain trajectories for 50+ frames without re-detection. Key innovations include (i) zone-adaptive background subtraction (16x16 grid, 6.1KB state) with dual learning rates, (ii) occupancy-aware updates to prevent stationary shoppers from background absorption, and (iii) oscillation-aware noise suppression to throttle unstable zones. CHURI achieves 78.2% recall on WEPDTOF overhead surveillance (+26.1pp over YOLO; +93% over MOG2), 87.6% on MERL Shopping, and 92.1% on Innovatiana Shoplifting, enabling Raspberry Pi 5 deployment at 1/70th traditional infrastructure cost.