Outdoor Motion Capture at Scale

Area of Science:

• Biomechanics
• Computer Vision
• Sports Science

Background:

• Capturing outdoor sports kinematics is difficult due to large volumes and environmental challenges.
• Current video-based motion capture relies heavily on manual post-processing, limiting its application.
• Automated detection of reference and sport-specific keypoints is needed to overcome these limitations.

Purpose of the Study:

• To develop and evaluate an automated motion capture pipeline for large-scale outdoor sports.
• To reduce manual post-processing time and improve the accuracy of kinematic data collection.
• To create sport-specific datasets for biomechanical research and AI model training.

Main Methods:

• A hybrid approach combining YOLO object detection and ArUco marker identification for reference point localization.
• Fine-tuning AlphaPose on a custom dataset for detecting skier-specific keypoints and anatomical landmarks.
• Utilizing Direct Linear Transformation for continuous frame-wise calibration and 3D reconstruction.

Main Results:

• Automated reference point detection achieved a mean localization error of 4.1 pixels.
• The skier-specific keypoint model reached 98% PCK, mAP of 0.97, and MPJPE of 10.3 pixels.
• Reduced 3D segment-length variation by 23% for reference points and 0.5-0.6 cm for skier keypoints compared to manual methods.

Conclusions:

• Automated detection significantly improves accuracy and efficiency in outdoor kinematic data collection.
• The pipeline facilitates large-scale data collection with multiple athletes and trials.
• The developed approach supports biomechanical research and the training of next-generation 3D pose estimation models.