Real-Time Running Performance Analysis and Optimization with YOLOv8-Based AI Vision and Random Forest Classification

Abstract

In this project, it was proposed to introduce an AI-enabled solution that identifies a running posture in real-time delivered through video analysis. The system employs the YOLOv8n-pose model to identify 17 anatomical keypoints in every video frame, which allows extracting three most important biomechanical properties: stride ratio, knee angle, and cadence. These characteristics are subsequently used to categorize every frame into Good or Not Good posture by a Random Forest classifier based on manually and semi-automatically labeled data. The model was also very accurate, as its precision had a value of 0.98, recall had 1.00, and F1- score was 0.99. Six running video tests were used to test the system and its results included CSV files, biomechanical graphs, and posture summaries. This study shows how non-invasive and low-cost solution can be applied instead of traditional motion capture that can be used in athletic training and rehabilitation tasks.