Model
ByteTrack
Assign persistent identities to detections across frames, surviving brief occlusions.
shipped
Multi-Object Tracking
TrackingMOT
Overview
ByteTrack associates both high- and low-confidence boxes to keep tracks alive through partial occlusion. Overseer layers a short coasting window on top: a last-known box is held briefly through a momentary detection drop so tracklets do not flicker. Track ids key everything downstream, Re-ID, speed, intent, ghosts and the tactical radar.
| Role | Multi-Object Tracking |
|---|---|
| Expected latency | < 2 ms / frame |
| Hardware requirements | CPU-bound; no GPU required |
| Memory requirements | Negligible |
Inputs & Outputs
Inputs
- Per-frame detections
- Previous track state
Outputs
- Stable track id per object
- Track age / velocity state
Advantages & Limitations
Advantages
- Robust through occlusion
- Cheap (association only)
- Foundation for all identity features
Limitations
- Identity switches possible in dense crowds
- Long full occlusions break continuity
- 2D only, no depth-aware association yet
Example outputs
ByteTrack
Consistent ids as subjects cross the frame
ByteTrack
A tracklet coasting through a 2-frame detection gap
Configuration
Keys in config/default.yaml that govern this model.
| Key | Default |
|---|---|
detectors.yolo.tracker | bytetrack |
tracking.coast_frames | 3 |
Benchmarks
Placeholder
Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.
| Benchmark | Dataset | Metric | Value | Hardware |
|---|---|---|---|---|
| MOTA | MOT17 (ref) | MOTA | ~80 | reference |
| Overhead | live | ms/frame | < 2 | CPU |