3D Pose Estimation and Tracking in Handball Actions Using a Monocular Camera

Area of Science:

• Computer Vision
• Sports Analytics
• Machine Learning

Background:

• Accurate player pose estimation is crucial for sports performance analysis, action recognition, and technique evaluation.
• Challenges include rapid movements, occlusions, and varying distances in team sports, demanding robust pose estimation and tracking.
• Current methods often struggle with real-world conditions, impacting the reliability of action recognition.

Purpose of the Study:

• To provide an overview and analysis of monocular camera-based player pose estimation and tracking methods for sports.
• To evaluate the applicability and robustness of deep learning methods for 3D pose estimation in handball scenarios.
• To propose and evaluate novel retargeting and smoothing techniques to enhance pose sequence accuracy.

Main Methods:

• Evaluated 12 two-stage deep learning methods for 3D pose estimation on handball jump shot datasets.
• Developed and tested retargeting and smoothing methods for 3D pose sequences.
• Assessed five state-of-the-art tracking methods on handball training data using monocular video.

Main Results:

• Proposed retargeting and smoothing methods experimentally improved performance across all tested 3D pose estimation models.
• Evaluated methods demonstrated varying degrees of applicability and robustness in real-world handball scenarios.
• Identified shortcomings in keypoint localization and pose generation, affecting overall action recognition accuracy.

Conclusions:

• Deep learning-based 3D pose estimation and tracking show promise but require further refinement for sports applications.
• The proposed pose sequence enhancement techniques offer a viable approach to improve accuracy.
• Addressing limitations in keypoint detection and anatomical plausibility is essential for future advancements in sports action recognition.