Multi-Camera Person Tracking and Re-Identification
Demo videos
Load the sample dual-camera videos to test tracking and cross-camera ReID.
Detector model
Tracker
0.10.5
Pipeline
Detection - YOLOv8 / YOLO11 pretrained weights detect people (COCO class 0) in each frame.
Tracking - ByteTrack or BoT-SORT links detections across frames, using a Kalman filter to keep IDs stable through short occlusions.
Re-identification - An OSNet backbone extracts L2-normalized 512-dimensional features from each person crop. Cosine distance between cameras is used to assign one global ID per person.
Interpolation - Short gaps in each track are filled by linear interpolation to reduce box flicker.