| Human motion capture data can accurately drive the movement of virtual people,so as to achieve a real and smooth movement effect.It has been widely used in television movies,3D games,virtual reality and other fields.With the massive growth of human motion capture data,how to manage and reuse these data has become an urgent problem to be solved.The analysis of existing human motion capture data has become the current research hotspot,which is important for the management and reuse of the current capture data.In this paper,the motion analysis of human motion capture data is carried out,including the behavior segmentation of motion capture data,single level semantic annotation and hierarchical semantic annotation.The main contents of this paper include:For the human motion capture data segmentation,this paper proposes a new segmentation method called Normalized Cut combined with Weighted Kernel K-means(NCWKK)method.We analysis the equivalence of matrix trace between the normalized cut model and the weighted kernel k-means,then do the clustering of the human motion capture data.The clustering results after the time sequence recovery are constructed a category string,and we define three different characters:valid substrings,long characters and invalid substrings to find out the segmentation points.The results of the experiments show that NCWKK method has a satisfactory segmentation performance.It can also find the segmented fragments of the same semantic types in human motion sequences.In the single level semantic annotation of human motion capture data,this paper proposes Derivative Piecewise Aggregate Approximation(DPAA)method for semantic annotation.We construct the relational feature functions by using the positional relationship between the human skeletal joints and the plane geometry,then obtain the single level motion templates.This method uses Piecewise Aggregate Approximation(PAA)to do the temporal reduction of motion sequences.By combining Dynamic Time Warping method with Derivative Dynamic Time Warping(DDTW)method,the sequences are sequentially matched with motion templates of different motion behaviors to achieve single level semantic annotation results.In the hierarchical semantic annotation of human motion capture data,this paper proposes hierarchical templates semantic annotation method.We construct Hidden Markov Model for motion sequences,then use the Expectation Maximization(EM)algorithm to calculate the parameters and hidden variables of the model,throughiteration to get the parent level motion category templates and the corresponding sub-level motion stylized templates.Then Kernel Canonical Time Warping(KCTW)method is used to do the feature space alignment operation of the parent-child motion templates and the unknown semantic sequences.Then we can obtain the corresponding warping distance,so as to achieve the annotating parent level motion categories and sub-level motion styles. |
