Human Joint Angle Estimation Using Deep Learning-Based Three-Dimensional Human Pose Estimation for Application in a...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomechanical Analysis

Background:

• Human Pose Estimation (HPE) methods, evolving from 2D to 3D, are vital for applications like AR, animation, and surveillance.
• Current 3D HPE methods struggle with real-world data due to limited training sets, depth ambiguity, left/right switching, and occlusions.

Purpose of the Study:

• To compare the performance of four 3D HPE methods using real-world videos.
• To propose and validate joint position correction techniques for improving 3D HPE accuracy in daily motions.
• To enable intuitive human activity recognition through corrected joint angle trajectories.

Main Methods:

• Comparative analysis of four 3D HPE algorithms on real-world video datasets.
• Development of joint position correction algorithms to address left/right inversion and false detections.
• Utilizing a 3D humanoid simulator for human activity recognition based on corrected joint angle trajectories.

Main Results:

• Identified strengths and weaknesses of different 3D HPE methods in practical scenarios.
• Demonstrated the effectiveness of proposed joint position correction in mitigating common HPE errors.
• Generated accurate joint angle trajectories for analyzing complex human movements.

Conclusions:

• The proposed joint position correction significantly enhances the reliability of 3D HPE for real-world applications.
• The method provides a robust approach for human activity recognition, particularly in dynamic scenarios like gymnastic exercises.