Applied AI · Computer Vision Project
Real-Time Human Pose Feedback
A real-time computer vision system that converts pose landmarks from live video into immediate movement feedback.
Context
Computer Vision Project
Role
Independent Project
Year
2026
Stack
Problem
What needed solving
Getting useful feedback on exercise form usually requires a trainer or a mirror and self-awareness most people do not have mid-movement; a camera-based system can watch continuously and flag form issues immediately.
What I built
Concrete responsibilities
- A capture loop that pulls frames from a live camera feed
- Pose landmark extraction using MediaPipe Pose
- Joint-angle and movement analysis computed from the tracked landmarks
- Real-time corrective feedback generated from that analysis
System
How it fits together
Each frame from the camera feed is passed to MediaPipe Pose to extract body landmarks. Joint angles and movement trajectories are computed from those landmarks and compared against target ranges to generate feedback with minimal latency.
Key technical decisions
What was interesting or non-trivial
Smoothed landmarks over a short rolling window instead of reacting to every frame.
Slightly less instantaneous feedback, in exchange for far fewer false corrections caused by single-frame tracking noise.
Result
What the system produced
A working real-time demo that tracks pose and surfaces feedback with low enough latency to be usable during movement, not just after it.
What I learned
Where the actual latency budget goes in a real-time vision system, and how much perceived responsiveness depends on smoothing choices rather than raw inference speed.