Low-Rank and Sparse Recovery of Human Gait Data.
Kaveh Kamali1, Ali Akbar Akbari2, Christian Desrosiers3
1Department of Automated Manufacturing Engineering, École de Technologie Supérieure, Montreal, QC H3C1K3, Canada.
Sensors (Basel, Switzerland)
|August 23, 2020
Summary
This study introduces novel unsupervised methods to reconstruct corrupted human motion data from optical tracking. The approach significantly improves accuracy, reducing errors by up to 14 mm compared to existing techniques.
Area of Science:
- Biomechanics
- Computer Vision
- Signal Processing
Background:
- Optical motion capture systems frequently lose data due to marker occlusion or detachment.
- Accurate human motion data is crucial for biomechanical analysis, rehabilitation, and animation.
Discussion:
- The study proposes two unsupervised reconstruction models leveraging low-rank matrix completion and group-sparsity priors in the frequency domain.
- These models utilize inherent properties of human gait to restore corrupted marker trajectories without requiring training data or user-specific kinematic information.
Key Insights:
- The novel methods demonstrate superior performance in recovering missing motion data across various gait datasets and gap lengths.
- A mean reconstruction error reduction of at least 2 mm was achieved compared to state-of-the-art principal component analysis (PCA) methods.
- When limited marker data is available, the proposed approach offers a significant improvement, reducing mean reconstruction error by over 14 mm.
Outlook:
- This unsupervised approach has the potential to enhance the reliability and accuracy of human motion capture in diverse applications.
- Future research could explore the integration of these techniques with real-time motion analysis systems.
- Further validation across different types of human movement and tracking scenarios is warranted.
