SP
All projects
archivedFeatured

Real-Time Yoga Pose Correction System

AI-powered mobile app using MoveNet for skeleton keypoint detection and a custom CNN for 90%+ accurate pose classification.

PythonTensorFlowMoveNetKerasAndroid StudioJavaTensorFlow.js

01Overview

This project leverages on-device computer vision to democratize personal yoga instruction. It achieved over 90% accuracy across multiple poses, providing low-latency corrective feedback directly on the device.

02Features

  • Real-time skeleton keypoint detection using MoveNet Lightning
  • Custom CNN achieving 90%+ accuracy in pose classification
  • Live corrective feedback system with visual overlays
  • Cross-platform support via TensorFlow.js for web browsers
  • Native Android implementation for low-latency performance

03Engineering Decisions

Chose MoveNet for its high performance on mobile hardware without requiring a cloud GPU.

Engineered the classification pipeline to be invariant to user distance and camera angle.

Integrated TensorFlow.js to enable a 'zero-install' web version alongside the native app.

04Challenges & Solutions

Challenge

Providing accurate feedback in varying lighting and environments.

Solution

Used robust keypoint normalization techniques to ensure the model focuses on joint relationships rather than absolute pixel positions.