TrafficVision: A Human-Verified Traffic Analytics and Number-Plate Annotation System for Pakistani Road Video

Pakistan’s road-safety planning is constrained by limited traffic data and road conditions that challenge conventional analytics systems, including handheld video, weak lane markings, heterogeneous vehicle types, and region-specific number plates. We present TrafficVision, a working human-in-the-loop AI system that transforms Pakistani road video into auditable traffic evidence. The system combines vehicle detection, persistent multi-object tracking, user-defined regions and tripwires, motion-aware geometry, helmet assessment with an explicit unknown state, and a workspace for human-verified number-plate annotation. Rather than equating frame-level detections with vehicle counts, TrafficVision records a vehicle only when a persistent track satisfies a defined geometric crossing event. Each reported event remains traceable to its source video, track, geometry, model version, and supporting frames. The current system snapshot includes 51 unique videos from 36 districts, 197,376 detections, 6,911 tracks, 858 deduplicated crossings, 394 reviewable safety events, and 115 plate annotations, of which 114 have been human-verified. These figures describe system activity and provenance rather than predictive accuracy. The demonstration presents the complete workflow, from video upload and motion-aware traffic analysis to event inspection and correction of machine-proposed plate annotations. TrafficVision illustrates how locally grounded, transparent, and human-supervised AI can support road-safety data development in resource-constrained settings without presenting automated predictions as enforcement-ready decisions.