Input
Predicted
84.1
MOTA
−63%
ID switches
11ms
Latency
Overview
A multi-object tracker designed to maintain identity across long occlusions in dense urban scenes. Built for real-time deployment in CCTV, robotics, and autonomous mobility contexts.
Method
We combine an appearance-based re-identification head with a motion-prior transformer that learns long-range temporal dependencies. The model can carry identity across occlusions of up to 30 seconds without losing track.
Deployment
Deployed in production on edge devices with 11ms end-to-end latency. Used by three logistics partners in real-world environments.