ADR-0005 — Leg IMU and/or camera CV for technical (football) metrics

Status: Accepted (Phase 3 direction) · Date: 2026-06-14

Context

A torso GPS+IMU vest gives physical metrics but cannot see shots/passes/dribbles — those are foot-to-ball contact events. The market reflects this: vest = physical, boot/calf = technical.

Decision

Defer technical metrics to Phase 3 and pursue one or both: - Path 1 — leg/boot IMU + ML classification (what Footbar/PlayerMaker do): a fast IMU (ICM-42688, ≥200–500 Hz; ball contact is ~5–15 ms) with gyro; one sensor per foot. Pipeline: peak detection → ±0.2 s window → features (peak |a|, peak |ω|, swing duration, energy, jerk, foot-in-air) → XGBoost/RandomForest → {shot, pass, touch, nothing}; later 1D-CNN/TCN. Shot power ≈ shank angular velocity × lever. Label via phone-video sync. - Path 2 — one elevated wide camera + open-source CV (Veo/Trace style): YOLO (players+ball) + tracking + homography → passes, possession, shots, positions for the whole team at once.

Recommended combination

Keep the GPS+IMU vest for live physical metrics + add one wide camera with an open-source CV pipeline for technical/ball metrics. The owner's AI/ML desktop (RTX 3060) makes the CV route viable. Open-source starting points: AmmarMohamed0/Football-Analysis-System (YOLO11 + ByteTrack), Khushal-gupta22/Football-Analysis (YOLOv8, team assignment via jersey-color K-means, ball interpolation, perspective transform, possession).

Consequences

• + Covers both metric families; camera gives the richest ball/possession data from one device.
• − IMU path needs labeled data and careful ball-strike vs ground-strike discrimination (gyro signature + foot-in-air). Camera path struggles with ball detection/occlusion and per-player attribution (jersey OCR / manual roster).
• Possession %, pass networks, and team shape are camera-only — not obtainable from wearables.

Alternatives considered

• Smart ball with 500 Hz IMU (Adidas connected ball) — rejected: can't DIY, and it tracks the ball, not individual players.