Science Tokyo research method improves person tracking in benchmarks

A person slips behind another person, an object or an obstruction. When they reappear in a different camera, the system may treat them as someone new. Researchers at the Institute of Science Tokyo, working with NEC Corporation, have developed a method that preserved cross-camera identities more reliably in benchmark tests, reaching an IDF1 score of 65.48 on MMPTrack.

The method combines two clues that answer different parts of the matching problem. Epipolar geometry describes the spatial relationship between two cameras and narrows down where the same person could appear. The system then compares visual appearance among the candidates that remain, using image features to distinguish people who are geometrically plausible matches.

The approach does not require replacing each camera’s existing single-camera tracking system. Those systems produce tracklets — short fragments of a person’s tracked movement — and the new method associates fragments across cameras. The researchers say it uses known camera relationships and pre-trained appearance features, so it does not require additional training for each environment.

The gains were measured against existing methods on two multi-camera benchmarks. On MMPTrack, the proposed system scored 65.48 IDF1, compared with 62.30 for MCTR, while its HOTA score — a measure combining detection accuracy and identity consistency — was 56.92, versus 55.77 for MCTR. On CAMPUS, it scored 47.37 IDF1, compared with 44.72 for ByteTrack.

And then, concretely? A system that keeps identities consistent across viewpoints could support security monitoring, transportation management, facility operations and pedestrian-flow analysis, without environment-specific retraining. But the result remains a laboratory-stage method: severe occlusion may prevent detection altogether, leaving no tracklet for cross-camera association. The research was presented at ICPR 2026 in Lyon and published in Lecture Notes in Computer Science.