qBoTSORT

A standalone BoT-SORT tracker built around a three-pass face-aware matching cascade, lifted out of qMetrics into its own repo and given a real 8-state Kalman filter, optional global motion compensation, and its own detectors and embedders.

The problem it solves: face evidence in the field is unreliable — a person crop can carry a clean face, no face at all (back turned, occluded), or several faces in a crowd. A tracker that collapses all of that into one appearance score mis-assigns constantly. qBoTSORT instead routes each detection through a different pass so bad face evidence never contaminates good body evidence:

• Pass 1 — gold multi-cue — confirmed tracks matched against clean single-face detections, scored on IoU + OSNet body re-ID + Buffalo face embedding + Kalman Mahalanobis distance
• Pass 2 — body-only — weak, absent, or multi-face detections; face is dropped and OSNet body re-ID carries the match (with a small bonus for the multi-face case)
• Pass 3 — motion-only recovery — lost tracks matched on IoU + Mahalanobis alone, tightly gated, echoing ByteTrack's second association stage

A retired-track gallery lets a person who ages out and later reappears reclaim their original id by face or body similarity rather than spawning a new one, and duplicate-track suppression drops the shorter-lived of two tracks sharing a box. It does tracking only — no demographics or zone logic layered on top.

Built on PyTorch, Ultralytics YOLO, InsightFace (buffalo_l), OSNet/torchreid, with S3 batch processing via boto3. Active research track, with recent work on a pose-aware re-ID gallery (using body-facing direction from pose keypoints) and a relative-stature veto for the matching cost.