Personal Project – First Pitch

Background

As a high school rowing athlete, I know that precise body movement is crucial in sports. In rowing, the catch angle of your arms, the drive through your legs, the sequencing of your hip hinge, and if any of these are even slightly off, you’re not just losing speed, you’re risking real injury. Overuse injuries from repeated improper form are incredibly common in sports, and the frustrating part is that most of these are preventable if you could actually see what your body is doing wrong.

But here’s the thing: it’s hard to quantify those movement errors into numbers. The human eye can’t detect a 5-degree difference in knee angle, but that 5 degrees might be the difference between staying health and getting injured.

Definition Statement

Current methods for analyzing human movement quality are either prohibitively expensive (professional motion capture systems like Vicon cost $150,000+) or too limited (consumer wearables only track basic metrics like steps and heart rate). This leaves everyday athletes, rehabilitation patients, and fitness enthusiasts without accessible tools to evaluate whether their movement form is biomechanically correct. The goal is to develop an affordable, AI-powered system that can perform full-body biomechanical motion analysis using only a standard camera, delivering expert-level feedback that was previously only available in clinical laboratories.

Who Am I Solving This For?

One of the most exciting things about this project is that the technology isn’t limited to one sport. The pose estimation tracks general human movement, and the scoring algorithm compares any two motion sequences. That means with the right reference recordings, the system can analyze a basketball free throw, a yoga pose, a deadlift, or a physical therapy exercise with the same accuracy.

There are over 15 million physically active Canadians and the global rehabilitation market is $287 billion, showing that the demand for accessible, accurate movement guidance is massive, but the supply is locked behind six-figure price tags or oversimplified apps that just count your reps.

My Solution: MotionX

My idea is to build an AI-powered movement analysis system called MotionX that uses a standard phone camera to perform real-time, full-body biomechanical analysis. Without using any markers, special hardware, expensive subscriptions. The system would:

1. Capture and track skeletal movement using MediaPipe’s pose estimation
2. Compare the user’s movement against a standard reference video using a DTW-based temporal comparison algorithm that handles natural speed variations
3. Score movement quality on a 0–100 scale with specific per-joint feedback
4. Provide AI-powered coaching through a language model that translates raw biomechanical data into plain language guidance and personalized training plans

The key technical differentiator is Dynamic Time Warping (DTW) — it doesn’t just check if you’re in the right position at one moment, it compares your entire movement trajectory over time against a professional reference, naturally handling the fact that you might move faster or slower. No consumer app currently does this.

I have previously worked on this idea a bit and developed a science fair version of MotionX that proved the concept works. But right now it’s a collection of individual parts: a Python desktop app, local video processing, API calls stitched together with scripts. It’s a prototype, not a product.

My main goal for this course is to take everything I’ve built and merge it into a single, polished, user-ready web application. That means building a real user system with account registration and login so people can track their progress over time. It means moving the motion analysis from a local Python script to an online platform where someone can upload or record video directly in their browser. It means setting up the AI coaching API properly. And it means designing an interface that anyone can pick up and use without reading documentation or understanding what DTW stands for.

What the Prototype Will Look Like

For testing and display by May, my prototype will be a functional web application with these core features:

• A webcam-based pose capture page that overlays the detected skeleton on the live video feed in real time
• A comparison engine that lets you record a movement, select a reference, and receive a DTW-based similarity score (0–100)
• A results dashboard showing per-joint angle comparisons (bar chart), an overall score, and specific correction suggestions
• AI-generated coaching feedback translating raw biomechanical data into actionable, plain-language advice

Step-by-Step Plan

Phase	Tasks	Deadline
Research & Planning	Finalize tech stack (React, Flask/Node, database), design system architecture, plan user flows and wireframes	Week 1
User System	Build account registration and login, set up secure database for user profiles, implement session management and password hashing	Week 2
Online Motion Analysis	Integrate MediaPipe pose estimation to run in-browser or via server, build video upload and recording interface, integrate DTW scoring pipeline with the web backend	Week 6
AI Coaching Integration	Set up AI API and connect three coaching agents (Professional Analysis, Training Plan, Expert Coach) to user-facing dashboard	Week 8
UI/UX Polish	Design clean, easy-to-use interface, add progress tracking and session history, build results dashboard with charts and joint-by-joint breakdowns	Week 10
Testing & Launch	User testing with real athletes and families, fix bugs, optimize performance, deploy and document	Week 12

Possible Problems

• Lighting and environment: Poor lighting can throw off pose detection. I’ll design a “setup check” screen to help users optimize their environment.
• Single-camera depth estimation: Estimating 3D positions from a 2D camera is inherently an approximation. Multi-camera setups would help but defeat the accessibility goal.
• Reference quality: The system is only as good as its reference movements. I’ll need to create high-quality reference standards, ideally from coaches or sports science resources.
• Browser performance: Running real-time pose detection in a browser may struggle on older hardware. I’ll need to test across devices and potentially offer a “lite” mode.
• DTW sensitivity plateau: Based on my research, Z-score normalization creates a scoring plateau at extreme deviations, and scores flatten around 89–90 even for very poor form. This is a known tradeoff that improves robustness but limits the scoring range.

Despite these challenges, the core technical feasibility has already been demonstrated in academic research. The opportunity is in the integration. Fusing pose estimation, temporal comparison, and AI coaching into a single accessible tool that nobody has built yet.