Research On The Key Techniques Of Reconstruction And Retrieve Based On Motion Capture Data

In recent years, with the rapid development of multimedia technology and computer graphics, motion capture technology has received wide attention and is successfully utilized in many real applications such as three dimensional animation, film production, and virtual reality. Accordingly, a large number of 3D human motion data has been gradually accumulated. Therefore, efficient motion data reconstruction, behavior recognition and motion retrieval are of crucial importance to the motion capture data reusing. In this thesis, we focus on studying the problems of motion denoising, incomplete motion recovery, motion recognition and semantic behavior retrieval. The main contents are organized as follows:1. Motion data denoising. We present a motion segmentation based human motion capture data recovery approach via the sparse and low-rank decomposition. First, Bilateral filter is adopted to reduce the noisy impacts and remove singular values,then probabilistic principal component analysis(PPCA) method is utilized to segment the motion data into different semantic behaviors automatically. Then, the accelerated proximal gradient(APG) algorithm is employed to achieve the partial data recovery and all the recovered sub-motions are sequentially combined to restore the whole motion. As a result, the real human behaviors can be well approximated.2. Incomplete motion data recovery. We develop two effective recovery approaches: 1) missing human motion capture data recovery via fuzzy clustering and projected proximal point algorithm(ProPPA); 2) hierarchical block-based incomplete human mocap data recovery approach by using adaptive nonnegative matrix factorization(ANMF). The former approach first utilizes linear interpolation method and fuzzy c-means(FCM) clustering algorithm to roughly fill the missing values and separate the motion sequence into several sub-clips. Then, ProPPA is employed to recover the sub-clip data partially and the recovered sub-motion clips are temporal recombined. Finally, the whole incomplete motion data can be well reconstructed. The latter approach represents the whole motion data as the block-based sub-chain motion clips and ANMF method is presented to restore each incomplete sub-chain motion clip individually by exploiting the low-rank structure and the nonnegativity synchronously. Finally, the whole incomplete mocap data can be well recovered by excavating the prior knowledge within the raw motion data. The experimental results have shown the outstanding performances.3. Motion recognition. We present an efficient motion capture behavior recognition approach via neighborhood preserving dictionary learning. First, all the motion sequences in the database are normalized to make the motion to be comparable. Then, the neighborhood correlations among the motion frames are exploited using Iterative Nearest Neighbors(INN) algorithm and subsequently added as the constraint for discriminative dictionary learning. Consequently, the behaviors can be efficiently recognized by sparse coding based classification scheme.4. Semantic motion retrieval. We address an efficient human motion retrieval approach via temporal adjacent bag of words(TA-BoW) and discriminative neighborhood preserving dictionary learning(DNP-DL). According to the simplified human skeleton model and the pairwise joint distances,we proposed a TA-BoW model to discriminatively code the motion appearances, in which the temporal properties within human motion are well exploited. Accordingly, the articulated complexity and spatiotemporal dimensionality could be greatly reduced. Subsequently, DNP-DL is exploited by considering the neighborhood relationships of intra-class structure and the advantage of Fisher criterion, whereby the motion behaviors can be sparsely represented by these discriminative dictionary atoms. Finally, we address a hierarchical retrieval mechanism by incorporating the sparse classification and Chi square ranking, whereby the searching range would be significantly reduced. The experiments have shown it promising performances.