Unified End-to-End YOLOv5-HR-TCM Framework for Automatic 2D/3D Human Pose Estimation for Real-Time Applications.

Area of Science:

• Computer Vision
• Machine Learning
• Biomechanical Analysis

Background:

• Three-dimensional human pose estimation is crucial for fields like sports, robotics, and healthcare.
• Existing Convolutional Neural Network (CNN)-based methods often prioritize accuracy over speed.
• A need exists for efficient and accurate end-to-end solutions for 3D human pose estimation.

Purpose of the Study:

• To propose a fast, unified, end-to-end model for 3D human pose estimation.
• To integrate best practices across person detection, 2D pose estimation, and 3D pose lifting.
• To develop a sports scoring application leveraging accurate 3D pose data.

Main Methods:

• The proposed YOLOv5-HR-TCM model utilizes a 2D-to-3D lifting approach.
• It combines optimized components for person detection, 2D pose estimation, and 3D pose reconstruction.
• The model was evaluated on the Human 3.6M dataset.

Main Results:

• The YOLOv5-HR-TCM model achieves high accuracy without sacrificing processing speed.
• The system processes at 3.146 FPS on a low-end computer.
• An average deviation angle of 8.2 degrees was achieved for sports scoring on the Human 3.6M dataset (Protocol #1).

Conclusions:

• YOLOv5-HR-TCM offers an efficient and accurate solution for 3D human pose estimation.
• The model's speed and accuracy enable practical applications, including automated sports analysis.
• The proposed method demonstrates the potential for real-time biomechanical analysis.