Cross-Platform Fitness App with On-Device Pose Detection

A fitness-tracking app that uses on-device machine learning to count reps and correct exercise form from the phone camera.

Doing exercises with poor form is a common cause of injury and slow progress, yet a personal trainer is not available to everyone. A phone camera and a pose-estimation model can watch a workout and give feedback. This project builds a cross-platform fitness app that counts repetitions and corrects form using on-device machine learning.

The app is built with Flutter, and a TensorFlow Lite pose-estimation model runs directly on the phone, so video never leaves the device, which protects privacy and works without internet. For each camera frame the model returns body landmarks such as shoulders, hips and knees. The app computes joint angles from them and uses simple rules or a small state machine to detect the stages of an exercise such as a squat or push-up, counting each complete repetition and flagging mistakes, for example knees caving in. Frame processing is optimized so that the display stays smooth on mid-range phones. Sessions, repetition counts and progress charts are saved and synchronized through Firebase so that users can follow their improvement over time. The project evaluates counting accuracy and speed on different devices.

You will learn on-device machine learning, real-time camera processing and cross-platform mobile development. The Project Reference Guide explains the algorithms and testing, and the Reference Implementation provides the app and model integration.