Fail-Safe Player Re-ID Tracking

I need a rock-solid, real-time player tracking module for football matches that guarantees the ID assigned to each athlete at kick-off never changes until the final whistle. Right now, our OpenCV–TensorFlow–YOLO pipeline sometimes swaps or loses IDs when athletes overlap, leave the frame briefly, or the camera angle shifts, and that ruins every speed, distance, position, and heat-map metric we generate.

Key requirements
• Sport: football.
• Camera setup: five or more synchronized feeds.
• Existing stack: OpenCV, TensorFlow, YOLO – your solution must plug into this environment.

What I expect
1. A multi-object tracker with integrated re-identification that preserves the same unique ID through occlusion, crossings, short disappearances, or camera changes.
2. Cross-camera association so the same player is recognised seamlessly when switching angles.
3. Sub-second latency suitable for live analytics.
4. An admin panel where I can select a player, watch them in real time, and review historical metrics and footage.
5. Clear API hooks or Python modules so we can pipe the stable IDs into our existing performance-analysis dashboards.

Acceptance test
• 90-minute match recording with deliberate occlusions and camera switches: zero ID swaps permitted.
• CPU/GPU usage and frame rate documentation for typical 1080p streams on an RTX-class GPU.

If you have production experience with deep-SORT, ByteTrack, FairMOT, or custom Siamese re-ID networks and can fine-tune them for football dynamics, let’s talk. Provide a brief outline of your proposed approach, relevant past work, and timeline for a functional prototype.