SportsIQ Architecture — Sports Science Meets Pose Estimation

Why MediaPipe over YOLO for Pose Estimation (Initially)

Kinetiq needs to extract meaningful biomechanical data from video — joint angles, symmetry scores, range of motion. The choice of pose estimation library determines what data you get and how fast.

MediaPipe: Fast, Accurate, Zero Setup

MediaPipe Pose gives 33 3D landmarks per person at 30+ FPS on CPU. No GPU required for single-person analysis. The z-coordinate (depth estimate) is a bonus — crude, but useful for calculating approximate joint angles in 3D space.

Why this matters for Sports Science: A squat analysis needs hip, knee, and ankle angles simultaneously. MediaPipe provides all three landmarks at consistent frame rate, making real-time feedback possible.

YOLO: Multi-Person, Team Contexts

For team sport analysis (basketball, football), you need to track multiple athletes simultaneously. MediaPipe degrades when people overlap. YOLO + pose estimation handles this better.

The planned architecture: YOLO for detection and tracking → MediaPipe for detailed per-athlete keypoint extraction.

pgvector for Movement Memory

Each session generates a movement embedding — a compressed representation of the athlete’s movement pattern. Storing these in PostgreSQL via pgvector enables:

• Session-to-session comparison (is the athlete improving?)
• Anomaly detection (this movement pattern is unusual for this athlete)
• Population comparison (how does this squat compare to 1000 other squat recordings?)

This is the part that turns a pose estimation tool into an actual sports science instrument.

Status

Architecture defined. Implementation starting.