The Study Of Human Motion Capture Data Sparse Modeling

With the sparse representation theory being more perfected, the sparse representation of the signal gains concern from more and more scholars. It has been used in various fields, and provides a good solution to numerous problems. Motion capture technology has become ever more mature. Analyzing and processing human motion capture data by using the sparse representation theory to obtain the better effect are the goal of this thesis. The work of this thesis mainly includes the following three aspects:(1) Sparse based semi-supervised distance learning for motion capture data similarity metric. Human motion retrieval and recognition technology are the essential issue for motion data management and reuse. Logically similar motions may be numerically dissimilar, so it is difficult to get feasible results if the logical similarity between the two movements is measured with Euclidean distance. This thesis presents a semi-supervised distance learning method for measuring the logical similarity with Mahalanobis Distance which is trained by labeled and unlabeled motion data. The experimental results show that the proposed method is effective for motion retrieval.(2) Key pose extraction is a fundamental function in various motion capture based animation technologies. In this thesis, we propose a novel model for key pose extraction by formulating the problem as an optimized sparse reconstruction model. The minimizer is expected to correspond to the extracted key poses. Instead of combining different types of features directly, the proposed model processes each type of feature separately, and utilizes novel consistent regularization to guarantee the consistency of the extracted key poses according to these different types of features. A triangle constraint is introduced to enforce that all poses are reconstructed only by their adjacent key poses. The experimental results demonstrate that the proposed method consistently outperforms state-of-the-art methods.(3) Since the parameters of existing parametric motion synthesis approaches are structurally inconsistent and have less intuitive meanings, we propose a sparse semantic parametric model. Our method automatically extracts several meaningful motion parameters which are able to control jumping height, walking path, and so on. Our model drastically reduces the operating complexity of motion synthesis and causes less lost in synthesis quality; by simply adjusting the values of such motion parameters, the generation of motion can be intuitively controlled, and natural motions can be synthesized in real-time.