Automatic motion capture data denoising via filtered subspace clustering and low rank matrix approximation

In this paper, we present an automatic Motion Capture (MoCap) data denoising approach via filtered subspace clustering and low rank matrix approximation. Within the proposed approach, we formulate the MoCap data denoising problem as a concatenation of piecewise motion matrix recovery problem. To this end, we first present a filtered subspace clustering approach to separate the noisy MoCap sequence into a group of disjoint piecewise motions, in which the moving trajectories of each piecewise motion always share the similar low dimensional subspace representation. Then, we employ the accelerated proximal gradient (APG) algorithm to find a complete low-rank matrix approximation to each noisy piecewise motion and further apply a moving average filter to smooth the moving trajectories between the connected motions. Finally, the whole noisy MoCap data can be automatically restored by a concatenation of all the recovered piecewise motions sequentially. The proposed approach does not need any physical information about the underling structure of MoCap data or require auxiliary data sets for training priors. The experimental results have shown an improved performance in comparison with the state-of-the-art competing approaches.

• We represent MoCap matrix by a mixture of multiple low-dimensional sub-matrices.
• We formulate the MoCap denoising as a concatenation of piecewise motion matrix recovery problem.
• We present a filtered subspace clustering to separate the noisy MoCap sequence.
• We introduce a label filtering scheme to regularize the subspace length.
• We find a complete low-rank matrix approximation to each noisy piecewise motion.