Computer Vision Meets Sports
Sports analytics has moved far beyond box scores. Modern teams and sports technology companies want frame-level insight: exactly where every athlete was, how their body moved, and what that means for performance and injury risk. That data has to come from somewhere — and the most scalable source is computer vision on the footage you already capture.
Hassan builds these systems end to end. At ID Sports Ventures in Berlin — first as a full-time Computer Vision Engineer (Sep 2025–Dec 2025), now part-time (Jan 2026–Present) — he works alongside sports scientists to turn multi-camera video into performance analytics used during live competition, not just in post-match reviews.
That design decision — live, not post-hoc — shapes everything: models must hit real-time latency on venue hardware, tracking must survive camera cuts and crowds, and the output has to be readable the moment a rally ends. It is product engineering with a latency meter attached, which is exactly the discipline most sports analytics deployments fail on.
Athlete Tracking Under Occlusion
Tracking athletes in a crowded frame is one of the hardest problems in applied vision. Players overlap, cross behind each other and re-enter the frame from odd angles — and a single identity switch corrupts the entire game's stats. Hassan solves this by combining YOLO detection with ByteTrack and DeepSORT, which maintain robust identity associations even through severe occlusion.
The tracking layer feeds a pose estimation stage, so every athlete has both a stable identity and a skeleton — the two data streams that make automated analysis possible. See what computer vision is if you're new to the space.
Biomechanical Performance Reports
Raw keypoints are just numbers. The value shows up when they become biomechanical metrics: joint angles, segment velocities, stance width, symmetry and movement quality. Hassan translates skeleton streams into these standardized measures and packages them as reports that coaches and scientists can act on without touching a line of code. The full methodology lives on the biomechanical analysis page.
Automated Scoring & Drill Assessment
Repetition-based sports — gymnastics, shooting, throwing, racket sports — run on drills that need objective scoring. Hassan builds pipelines that automatically score technique by comparing pose trajectories against target form, flagging deviations in real time. Athletes get instant feedback; coaches get consistency across every attempt, not just the ones they happened to watch closely.
Scoring is deliberately transparent. Every automated score comes with the underlying keypoint evidence, so a coach can see why a rep was flagged — which keeps the system credible with athletes and useful in training, rather than a black box that gets ignored.
Proven Results
Measurable outcomes from shipped work:
- +15% pose accuracy — through data curation and preprocessing improvements, not bigger models.
- −40% preprocessing overhead — headroom reinvested into lower latency and richer analytics.
- Sub-30ms inference — TensorRT-accelerated models running live on edge hardware.
Dig deeper into the domain through the sports analytics overview, and if you're building a similar product, the engineering approach applies to teams, leagues and training facilities alike.
“The goal isn't more data — it's a coach who can say 'that's a problem' five seconds after it happens, backed by numbers.” — Muhammad Hassan Gul