VisionTrack

Real-time person recognition and tracking for smart surveillance — sub-100ms across live camera feeds.

Computer Vision at the Edge

VisionTrack is a smart surveillance platform providing real-time person recognition and tracking across distributed camera networks — built for a retail chain to reduce shrinkage and optimize staff deployment.

VISIONTRACK

Challenge & Solution

The Challenge

A retail chain needed real-time person detection and re-identification across 80+ cameras in multiple stores — with sub-100ms latency and the ability to flag watchlist matches without cloud dependency.

Our Solution

An edge-first computer vision pipeline using InsightFace for face detection and recognition, pgvector for similarity search against a managed watchlist, and Redis for sub-100ms cross-camera correlation.

Key Features

Real-time face detection and recognition across 80+ concurrent camera feeds
pgvector similarity search for watchlist matching at sub-50ms latency
Edge inference on NVIDIA Jetson hardware with cloud sync for analytics
Cross-camera person re-identification and movement path tracking
Configurable alert thresholds with Telegram and email notifications
Privacy-compliant design with on-device encoding and no raw image retention

Results & Impact

<100ms

End-to-end recognition latency

80+

Camera feeds processed

97.3%

Recognition accuracy

41%

Shrinkage reduction

Start a Similar Project