Pose EstimationYOLOComputer VisionStreamlit
Rider AI — MTB Performance Analysis
Computer-vision pipeline for MTB downhill analysis. Detects rider pose frame-by-frame, extracts biomechanical metrics, and generates AI coaching feedback via Groq LLM.
The Problem
Mountain bikers lack objective, data-driven feedback on their technique. Coaching is scarce, expensive, and subjective — making it nearly impossible to identify specific posture or line-selection issues from raw footage.
The Solution
A real-time computer vision pipeline that processes MTB video, detects 17 COCO keypoints per frame with YOLO26 Pose, extracts biomechanical metrics (balance, trunk angle, knee/elbow flexion), and generates personalized coaching feedback via Groq LLM.
Architecture
01YOLO26 Pose — 17-keypoint extraction per frame
02Moving-average keypoint smoother for stable skeleton
03Per-frame feature builder: balance score, trunk angle, posture class
04Trajectory and terrain type classifier
05Groq LLM coaching integration (OpenAI-compatible)
06Streamlit interface with Gallery and annotated video export
Results & Outcomes
Real-time pose estimation at 30+ FPS on standard hardware
Attack vs. defensive posture classified per frame
H.264 annotated video output with skeleton overlay and live stats
Deployed publicly on Streamlit Cloud — zero setup for end users