Research On Sparse Representation Based Human Motion Reconstruction

Recently, motion capture technique is becoming more and more popular and widely adopted, thus producing a lot of3D human motion data. Meanwhile, there are two appealing questions about motion data. Firstly, even the commercial mocap system cannot avoid noise, due to marker occlusions or mislabeling. In order to use accurate motion data in industry, raw data must be denoised and manual correction by professional animators is usually very time-consuming. Secondly, existing commercial motion capture systems require performer to wear cumbersome equipment, such as forty retro-reflective markers attached to human body joints. However, in applications for home use, users only care some key joints’position and don’t care the accuracy of other joints.Therefore, we propose human motion denoising method and human motion reconstruction method from a few joints’information respectively.This paper proposes a data-driven approach to remove motion noise based on sparse representation. Firstly, we construct a more fine-grained human spatial-temporal feature. For input motion with noise, we extract this feature and search k nearest neighbors to calculate a local representable dictionary from database. Then we handle the motion denoising problem using sparse representation optimization framework. For two most common noises, i.e. Gaussian noise and outliers, we adopt two slightly different objective functions to remove them respectively.This paper proposes a human motion reconstruction method from a few key joints’ information. For each input frame, we search k nearest poses which are most compatible with the control signals. These k nearest neighbors constitute local dictionary. Finally sparse representation optimization framework is used to reconstruct the body pose with all joints.