Computer Vision Engineer — Player Tracking & Sports Video Analytics (YOLO, ByteTrack)

We're building a computer-vision pipeline that turns raw soccer match video into player tracking data, physical/tactical performance metrics, and opponent scouting profiles, inside a Django-based sports SaaS platform. We need a computer vision engineer to take this from an early prototype to a validated, production-quality pipeline, delivered in sequential phases. CURRENT STATE Player detection runs on YOLO but without proper filtering/validation; tracking is a rough prototype with no occlusion handling; team assignment, real-world coordinate mapping, physical metrics, ball detection, and event detection are either missing or not reliable. SCOPE — 6 PHASES (each independently deployable behind a config flag) - Phase 0: fix detection filtering, real FPS handling, and build an evaluation harness (annotated reference clips + automated scoring) that every later phase is measured against. - Phase 1: integrate a robust multi-object tracker (e.g. ByteTrack) and assign teams via jersey-color clustering with temporal voting. - Phase 2: camera-to-pitch coordinate calibration (homography) so positions can be measured in real meters. - Phase 3: trajectory smoothing and real distance/speed/sprint/acceleration metrics with plausibility checks; remove fabricated team metrics, add the ones genuinely computable (compactness, width, etc.). - Phase 4: ball detection (small, hard object), possession, passes, duels, shots. Highest-uncertainty phase — we plan to start with a time-boxed proof of concept before committing to full scope. - Phase 5: population-based normalization of player/team profiles. - Phase 6: production hardening — adaptive performance tuning, GPU evaluation, QA artifacts visible to end users. You don't need to commit to all 6 phases up front — happy to start with Phase 0–1 (or 0–2) as a paid trial engagement and continue from there. TECH ENVIRONMENT Python, Django, OpenCV, Ultralytics YOLO, Supervision (ByteTrack), SciPy, scikit-learn. Heavy CV code runs inside a Celery worker, isolated from the Django serverless deployment. WHAT WE'RE LOOKING FOR - Proven experience with object detection (YOLO or similar) and multi-object tracking in video. - Experience with camera calibration / homography for pixel-to-real-world coordinate mapping. - Comfortable defining measurable, defensible metrics — validated against ground truth, not "looks about right." - Bonus: sports analytics, small-object detection (ball/puck tracking), or SAHI/tiling techniques. - Python proficiency; able to work inside an existing Django/Celery codebase without introducing heavy imports into the serverless layer. A short technical brief is available attached to this post.