Human Motion Analysis: Tracking Athletic Performance, Injury Prevention, and Movement Disabilities

Preview:

Contents:

• Overview
• Pipeline
• Example Input --> Output
• Outputted Metrics
  • Dashboard Video
  • Phase Overlay Videos
  • Feature Deviation Tracking
  • Total Form Deviation
  • Phase Breakdown Isolation
  • Phase Deviation Tracking
  • Ground Contact Stats

Overview:

Current Capabilities:

This project aims to create an inexpensive and accessible system for analyzing human movement from video.

Using pose estimation and machine learning, the system extracts biomechanical features from videos of users sprinting to evaluate performance, detect movement patterns, and provide feedback suggestions.

---

Future Implementations:

The long-term goal is to support applications in:

• Other athletic actions (baseball swing, barbell squat, basketball jumpshot, etc.)
• Injury prevention feedback
• Rehabilitation
• Motor Planning
• Movement Disorder Assessment

---

Project Motivations:

As a track athlete, I wanted to create a tool that people can use to evalutate sprint performance without access to expensive lab equipment.

I got the idea to expand the capabilites of the tool to help enhance motor planning because of my younger brother with special needs.

---

Pipeline:

Step	Goal	Tool	Use
1	Pose Estimation	MMPose, Yolo26, Mediapipe	Extract pose landmarks, joint angles, velocities using the user's choice of an engine for pose estimation between: Google's MediaPipe, Ultralytics YOLO26, or OpenMMLab's MMPose libraries, with custom normalization to enable standard comparison across datasets.
2	Phase Identification	Keras 1D Convolutional Neural Network	Predict gait phases using a 1D Convolutional Neural Network that analyzes temporal patterns in user landmark sequences. Phases are separated into: - Left/Right Ground Contact (LGC/RGC) - Left/Right Propulsion (LP/RP) - Left/Right Flight (LF/RF)
3	Form Scoring	Median Absolute Deviation (MAD)	Compute deviation from reference motion patterns using Median Absolute Deviation (MAD) and calculate each feature's similarity scores over time using MAD-based Z-scores.
4	Output	OpenCV, Matplotlib, Plotly	Generate visualizations: OpenCV: - Dashboard Video containing Form Deviation over time, color coded skeleton, and form correction suggestions - Phase Overlays of Correct Form Matplotlib: - Individual Z-scores over time - Total Form Deviation Scoring over time - Phase Z-scores over time Plotly: - Individual Phase Breakdown Isolation

Example:

Input & Skeleton Overlay user_input.mp4

Biomechanical Dashboard (Annotated) annotated_dashboard.mp4

NOTICE:

In dashboard video (pictured right), the large spike in Total Form Deviation at ~11sec aligns with form error of athletes left arm being raised This validates the ability of the pipeline to identify abnormalities in running form.

Outputted Metrics:

Full Video Analysis:

• Dashboard Video
• Feature Deviation Tracking
• Total Form Deviation Tracking

Individual Phase Analysis:

• Phase Overlay Videos
• Phase Deviation Tracking
• Phase Breakdown Tracking
• Phase Breakdown Isolation

Ground Contact Statistics:

• Ground Contact Times
• Left / Right Imbalance
• Strike Point Statistics

Example Outputs:

Metric	Example Output
Dashboard Video	dashboard.mp4
Phase Overlay Videos
Feature Deviation Tracking
Total Form Deviation
Phase Deviation Tracking
Phase Breakdown Isolation
Ground Contact Stats	----------------STRIDE FREQUENCY---------------- 202.068 Steps per Minute --------------GROUND STRIKE POINTS-------------- Average Right Ground Strike Point: 0.6188 Average Left Ground Strike Point: 0.5970 Average Strike Point Imbalance: (negative = left | positive = right) 0.0217 --------------GROUND CONTACT TIMES-------------- Average Ground Contact Time: Frames: 6 Seconds: 0.195 Average Right Ground Contact Time: Frames: 6 Seconds: 0.193 Average Left Ground Contact Time: Frames: 6 Seconds: 0.197