Player tracking pipeline for football and padel footage (fixed price, open source stack)

Building a sports analytics platform that delivers personal stat cards to amateur football (5 and 7-a-side) and padel players via WhatsApp, within minutes of their match ending. A fixed camera records the match; the pipeline tracks every player and outputs per-player stats. No wearables, no apps. What I need you to build: the CV pipeline engine Input: MP4/MOV, 1080p, 30fps, 60 to 90 minutes (padel: 1 camera behind the back glass at ~4m height; football: 2 cameras behind the goal lines at 6 to 8m). Plus a JSON of first-frame player bounding boxes mapped to names (provided by a separate tagging tool). Output per player: positions in metres over time, total distance, zone breakdowns, heatmap grid (for football: thirds; for padel: net vs baseline), sprint count. Plus: timestamped events (goals, shots for football; net approaches for padel) each with a confidence score. Plus: an annotated QA video. Requirements Stack: Python, supervision + roboflow/sports + ByteTrack. Detector must be swappable; RF-DETR (Apache 2.0) preferred. No Ultralytics YOLO in production (AGPL licence issue). Homography: map player foot points to real-world metres using pitch/court keypoints or a 4-point manual calibration per camera (one-time per venue). Ball tracking: dedicated model or SAHI tiling with interpolation. Runs on Modal (serverless GPU). A 60-minute match must process in under 15 minutes using parallel 1-minute chunks with overlap. Designed for live streaming: 1-minute chunks arrive in real time; tracks must stitch across chunk boundaries. Full IP assignment. Documented code. No dependency on any AGPL-licensed library in production. Milestones: Paid trial (48h): run your existing pipeline on 5 minutes of footage and share results Player tracking + homography + heatmap/distance JSON output Ball tracking + events with confidence scores Modal deploy, chunk stitching, documentation, handover Acceptance criteria: Identity accuracy ≥90% on a 10-minute sample. Heatmap matches the footage visually. 60-minute match processes in under 15 minutes on Modal. Results reported against our gold test set. You’re a strong fit if you have: - A sports tracking video you built (not a tutorial follow-along) - Experience with homography or camera calibration in a real project - Worked with supervision, ByteTrack or RF-DETR - Can explain how you prevent ID swaps across occlusions