Problem
A video analytics team computes statistics such as people counts from camera streams, and almost every module depends on a tracker. A standard tracking-by-detection pipeline (detector plus tracker) assigns a new ID when a person is occluded by another person or leaves the frame and returns. Each switch breaks the link between a person and an ID, which the statistics rely on. The pipeline has to run online, at 10 FPS or more.
Example from the article: in the CubicSiri video analytics platform, a person who disappears from the camera view for a while should get their previous identifier back when they return.
Solution in brief
Add a separate re-identification step in front of the usual association stages. Tracks that the motion model has lost but not yet deleted are matched with active tracks by cosine distance between ReID embeddings. Only "pure" embeddings, taken when a person's box does not overlap other boxes or the frame edge, are used for this.
Questions this case answers
- How to keep stable identities of moving objects across occlusions and re-entries in online tracking when the motion model has lost them but they are not yet deleted
- How to add an appearance-based matching stage to a tracking-by-detection pipeline without retraining the detector or the embedding model, at real-time frame rates
- How to keep stable person IDs across occlusions when counting people from video in real time
- How to return the previous ID to a person who left the camera view and came back
Tools and methods
- ByteTrack — the motion tracker the team built on (two-stage IoU association with a Kalman filter).
- ReID embeddings (OSNet, 2.2 million parameters) — vectors of person crops, compared by cosine distance.
- PureTrack — the team's tracker: independent ReID merge, pure-embedding filter, new "re-identifiable" track state.
- BoxMOT — library for running and scoring trackers on MOT17.
- MOT17, HOTA, MOTA, IDF1 — benchmark and metrics used for validation.
Google or a one-time email code. No password.