Human Motion Enhancement via Tobit Kalman Filter-Assisted Autoencoder

Area of Science:

• Computer Vision
• Biomechanical Engineering
• Machine Learning

Background:

• Low-cost depth sensors for human motion capture (D-Mocap) suffer from inaccuracies due to occlusion, interference, and algorithmic limits.
• Existing methods lack robust solutions for improving D-Mocap data quality.
• A need exists for reliable and accurate human motion data from affordable sensors.

Purpose of the Study:

• To develop a novel approach for enhancing the quality of human motion data captured by low-cost depth sensors.
• To improve the accuracy and stability of D-Mocap data.
• To introduce a new benchmark dataset for D-Mocap research.

Main Methods:

• Learning a general motion manifold using a convolutional autoencoder with diverse Mocap data.
• Incorporating the Tobit Kalman filter (TKF) for kinematic capture and censored data handling.
• Integrating TKF with the autoencoder via latent space optimization for manifold adherence and kinematic preservation.

Main Results:

• The proposed algorithm significantly improves the accuracy of joint positions and angles.
• Skeletal bone length accuracy is enhanced by over 50%.
• Experimental results demonstrate the effectiveness of the approach across simulated and real-world D-Mocap datasets.

Conclusions:

• The novel approach effectively enhances D-Mocap data quality, achieving over 50% improvement in accuracy.
• The developed extended MHAD dataset provides a valuable open-source benchmark for future research.
• This work advances the field of affordable and accurate human motion capture.