Optically Non-Contact Cross-Country Skiing Action Recognition Based on Key-Point Collaborative Estimation and Motion...

Area of Science:

• Sports Science
• Computer Vision
• Biomechanics

Background:

• Motion recognition in cross-country skiing aids technique optimization and strategy development.
• Non-contact visual sensors offer potential for ski training but face challenges from athlete posture changes, environmental variability, and limited fields of view.
• Existing methods struggle with the complexities of real-world ski environments.

Purpose of the Study:

• To enhance the applicability of monocular optical sensor-based motion recognition for cross-country skiing.
• To develop a robust method for detecting athlete posture and recognizing skiing movements using a single camera.

Main Methods:

• A monocular posture detection method utilizing cooperative detection and feature extraction was proposed.
• The method employs four feature layers for simultaneous human posture and key point detection.
• A loss function incorporating position deviation and rotation compensation was used for 3D key point estimation, followed by feature extraction for movement recognition.

Main Results:

• The proposed method achieved 90% accuracy in recognizing cross-country skiing movements.
• Performance was comparable to recognition methods employing wearable sensors.
• The algorithm demonstrated effectiveness in identifying typical cross-country skiing movement stages and sub-movements.

Conclusions:

• The developed monocular vision algorithm offers a viable, non-contact solution for cross-country ski motion recognition.
• This method holds significant application value for scientific training and performance analysis in cross-country skiing.
• The approach addresses key challenges in applying visual sensors to dynamic sports environments.