Deep-Learning-Based Recovery of Missing Optical Marker Trajectories in 3D Motion Capture Systems

Area of Science:

• Biomechanics
• Motion Analysis
• Deep Learning

Background:

• Motion capture (MoCap) data loss due to occlusions and technical issues is a significant challenge in biomechanics.
• Traditional data recovery methods have limitations in accuracy and robustness.

Purpose of the Study:

• To introduce a novel deep learning technique for recovering missing MoCap data.
• To enhance the accuracy and robustness of MoCap data reconstruction, especially for long-term data loss.

Main Methods:

• Developed a U-net-inspired bi-directional long short-term memory (U-Bi-LSTM) autoencoder.
• Utilized multi-camera and triangulated 3D data with a U-shaped deep learning structure.
• Incorporated an adaptive Huber regression layer for outlier robustness.

Main Results:

• The U-Bi-LSTM method demonstrated statistically significant improvements in reconstruction error compared to traditional methods.
• Achieved superior performance across various data gap lengths and numbers.
• Showcased effectiveness in long-term data loss scenarios.

Conclusions:

• The novel U-Bi-LSTM approach offers a robust and accurate solution for MoCap data recovery.
• This advancement enriches analytical tools for biomechanical research, impacting athletic performance and rehabilitation.
• Enables more precise biomechanical data for personalized treatment plans.