Multi-Object Detection & Persistent ID Tracking

Upload a sports video to detect and track players with persistent unique IDs. Uses YOLOv11 for detection and BoT-SORT (with ReID + camera motion compensation) for tracking.

How it works: Each detected person gets a unique ID that persists through occlusions, camera pans, and similar appearances thanks to appearance embeddings and motion compensation.

Upload Video

Detection Model

n=fast, m=balanced, l=accurate

Tracker

ByteTrack (install boxmot for BoT-SORT)

Detection Confidence Threshold

0.1 0.9

Annotated Output Video

Movement Heatmap

Object Count Over Time

Technical Details:

Detection: YOLOv11 (COCO-pretrained, class 0 = person)
Tracking: BoT-SORT with OSNet ReID embeddings + sparse optical flow camera motion compensation
ID Persistence: 5-layer strategy — Kalman filter, IoU association, ReID embeddings, camera motion compensation, track lifecycle management

GitHub Repository | Technical Report