TL;DR
This paper introduces two novel, unsupervised methods for reconstructing missing human gait data from optical motion capture systems. By leveraging the inherent low-rank and sparse properties of gait, their approach significantly outperforms existing methods like PCA, especially when dealing with limited marker data.
What Problem Does This Paper Address?
Capturing human motion with optical tracking systems, particularly for gait analysis, is prone to data loss. This can occur due to occlusions (markers being hidden by body parts) or system errors, leading to corrupted or missing data points. This missing data is a significant challenge for accurate gait analysis, which has applications in fields ranging from medical rehabilitation and prosthetics to animation and robotics. Traditional methods like linear interpolation struggle with longer gaps, while other techniques may require extensive training data or specific kinematic information, which isn't always available. This paper aims to address the problem of reconstructing these missing gait trajectories in an unsupervised manner, without needing prior knowledge of the user's specific movements.
Methods
The authors propose two related methods, both based on the principles of low-rank matrix completion and sparse recovery. They exploit three key properties of human gait:
- High Correlation: Trajectories of different markers on the body are highly correlated. This suggests the data matrix can be represented by a low-rank structure.
- Periodicity and Low Frequency: Gait is periodic and primarily composed of low-frequency signals, meaning its representation in the frequency domain is sparse.
- Joint Sparsity: Different limb trajectories share common frequencies, indicating a joint sparsity in their frequency domain representations.
The core of their approach involves formulating an optimization problem that seeks to find a complete data matrix X that is both low-rank and sparse in its frequency domain representation, while also matching the observed data. They use the Alternating Direction Method of Multipliers (ADMM) algorithm to efficiently solve this complex optimization problem.
The two proposed methods differ in how they model the sparsity:
- Sparse Low-Rank (S-LR): Uses an L1-norm to enforce sparsity on the frequency-transformed data.
- Group-Sparse Low-Rank (GS-LR): Employs an L1,2-norm to model the joint sparsity of frequencies across different trajectories, leveraging the shared frequency components.
Datasets
The authors evaluated their methods on four gait datasets:
- SlowWalk, FreeWalk, FastWalk: These were collected by the authors on a single healthy adult male at different speeds.
- WalkL: A publicly available dataset from a previous study [31], consisting of 10 consecutive walking cycles.
For evaluation, they simulated missing data by introducing synthetic gaps of varying lengths (10-50% of a gait cycle) across markers and time points. They tested their methods in two settings: using all available marker trajectories (37 markers) and using a reduced set of marker trajectories (11 markers, one per limb).
Key Insights and Ideas
- Gait as a Low-Rank and Sparse Signal: The paper provides strong empirical evidence that human gait data, when viewed as a matrix of marker trajectories, inherently possesses low-rank and sparse properties, particularly in its frequency domain representation. This is a fundamental insight that underpins their entire approach.
- Frequency Domain Sparsity is Key: By transforming the data into the frequency domain, the authors can effectively exploit the sparsity of gait signals. This is a more powerful approach than just relying on temporal sparsity.
- Joint Sparsity for Enhanced Recovery: The GS-LR method's ability to model the joint sparsity of frequencies across different limbs offers a significant advantage, leading to better reconstruction accuracy, especially when dealing with a full set of markers.
- Unsupervised and Data-Agnostic: A major strength is that their methods are fully unsupervised and do not require any prior training data or knowledge of the user's specific biomechanics. This makes them highly versatile.
- Robustness to Limited Data: The proposed methods demonstrate remarkable robustness when only a small number of marker trajectories are available, significantly outperforming existing methods in such scenarios.
Key Results
- Outperforms PCA: Both S-LR and GS-LR methods consistently outperformed Principal Component Analysis (PCA), a common baseline, across all tested datasets and settings.
- Significant Error Reduction:
- With all 37 markers, GS-LR achieved a mean reconstruction error of 8.27 mm on the WalkL dataset, compared to 12.5 mm for PCA.
- With only 11 markers (one per limb), GS-LR achieved a mean error of 12.7 mm, while PCA resulted in 28.5 mm. This represents a reduction of over 14 mm.
- GS-LR Superiority: Statistical analysis (Wilcoxon signed rank test) indicated that GS-LR generally achieved smaller errors than S-LR when all marker trajectories were used (p < 0.05).
- Impact of Sparse Regularization: The authors showed that the sparse regularization parameter (λ) has a clear impact on performance, with optimal values found between 10⁴ and 10⁵.
Limitations and Future Work
- Limited Population Size: The study acknowledges that the generalization of findings to a wider population is limited due to the small sample size used for data collection.
- Training Data for Improved Accuracy: While their current methods are unsupervised, they note that incorporating training gait datasets could potentially provide even better reconstruction accuracy, though this would introduce time-consuming training processes and require large datasets.
- Future Directions: The authors suggest extending their methods to capture data from different populations (e.g., elderly individuals, those with gait disorders) and integrating other data sources like foot pressure patterns and electromyographic (EMG) data. They also propose exploring other technologies and methodologies for more comprehensive gait analysis.
Summary updated