Recognition and Scoring Physical Exercises via Temporal and Relative Analysis of Skeleton Nodes Extracted from the...

Area of Science:

• Computer Science
• Biomedical Engineering
• Sports Science

Background:

• Human activity recognition is crucial for interactive systems like computer games.
• Current vision-based and depth sensor solutions for activity recognition face unresolved challenges.
• Kinect sensors offer a depth-sensing approach for capturing human motion data.

Purpose of the Study:

• To develop a pattern recognition-based solution for labeling and scoring physical exercises using Kinect sensor data.
• To quantify relationships between human skeletal joints for accurate exercise motion analysis.
• To achieve high accuracy in both exercise identification and performance scoring.

Main Methods:

• Feature extraction from human skeletal joints to generate relative descriptors.
• Utilizing discriminating descriptors to identify adaptive kernels in the Constrained Energy Minimization method for motion detection.
• Employing a geometric method for interpolating descriptor vectors to derive semantic exercise scores.

Main Results:

• Achieved 95.9% accuracy in the labeling process of physical exercise motions.
• Demonstrated a 99.5% agreement (R² index) between automated scoring and sports coach evaluations.
• Quantified meaningful relationships between skeletal joints during exercise performance.

Conclusions:

• The proposed pattern recognition method effectively labels and scores physical exercises captured by Kinect sensors.
• The system provides a reliable and accurate tool for automated exercise analysis.
• This approach has significant potential for applications in fitness, rehabilitation, and interactive gaming.