Mining Spatial-Temporal Patterns and Structural Sparsity for Human Motion Data Denoising

Area of Science:

• Computer Vision and Animation
• Biomechanical Engineering
• Data Science

Background:

• Motion capture is crucial for various fields but suffers from noise and outliers in acquired data.
• Existing denoising methods struggle with the complexity and variability of human motion.
• Addressing noise in motion capture data remains a significant challenge.

Purpose of the Study:

• To develop a robust, data-driven method for denoising human motion capture data.
• To improve the accuracy and stability of processed motion data.
• To offer a more accessible training dataset setup compared to existing methods.

Main Methods:

• Utilized a fine-grained partlet model for feature representation, capturing local body part similarities.
• Employed a robust dictionary learning algorithm to create multiple representative motion dictionaries.
• Formulated denoising as a structured sparse coding problem, integrating noise distribution and temporal smoothness.

Main Results:

• The proposed method demonstrated superior performance over state-of-the-art techniques on both synthetic and real-world noisy data.
• Outputs from the new approach exhibited enhanced stability.
• The training dataset setup for this method is notably simpler.

Conclusions:

• The novel approach effectively denoises human motion data by leveraging spatial-temporal patterns and structural sparsity.
• This method offers a more stable and accurate solution for motion capture data processing.
• The improved ease of training dataset creation makes it a practical advancement.