Research On Virtual Human Motion Synthesis Techniques And Engineering Application

The motion synthesis technology based on motion capture(mocap) data can create natural, realistic and exquisite human motions for virtual human characters, and it can greatly reduces the workload and work intensity for producing the animation, so it is of great significance for"people-centered design"in the field of production designing.This dissertation deeply researched on the virtual human motion synthesis based on mocap data, including motion primitive segmentation of virtual human, motion retrieval technology, motion synthesis and motion editing technology. And the paper also discussed the application of mocap-based motion synthesis in the field of engineering. The main innovations and achievements of this paper are listed as follows:(1) Aiming at the traits of mocap data, a new method of mocap data segmentation based on kernel dynamic texture is proposed. On the observation of the similarity between the mocap data and the dynamic textures, the dissertation introduced the dynamic texture model into the field of mocap data analysis. Through modeling mocap data by kernel dynamic texture(KDT), and using Martin distance as similarity measurement of different KDTs, we first segment motion sequence into sub-sequences with different motion style, then we segment the sub-sequence into cycle sequences and transition sequences, finally, for cycle motion sequence, if it is symmetrical, we segment it into two symmetrical motion clips, and thus, we solve the motion segmentation problem through a multi-step scheme, in which motion sequences are segmented into motion primitives from coarse to fine. Experimental results show that for periodic motion data, our approach provides the good performance, and the accuracy can be about 90%.(2) In order to improve the efficiency and accuracy of motion retrieval, a motion retrieval method based on Hidden Markov Model(HMM) clustering is proposed. Human motion data is time-series which has a certain probability distribution, and the length of the segmented motion primitives as well as motion clips to be retrieved are generally different, making the the retrieval difficult. On the observation of the strong modeling capabilities of time-series data of HMM, this paper implemented the K-means clustering method based on HMM to cluster motion primitives in motion database.Experimental results show that the accuracy of the clustering method can be above 95%. In the retrieval stage, by using HMM observation probability as similarity measurement between motion primitives or motion primitive classes, we first get the motion primitive class which has the similar motion style with the retrieved motion clip, then, by HMM observation probability, the most similar motion primitive in the class can be retrieved. Thus, we implemented the motion retrieval method. Experiment results show that our approaches are effective for motion data retrieval.(3) By synthesizing a sequence of complex motion in a virtual environment, we validated the whole motion synthesis techniques proposed in the paper, and by this example, we researched on the application of motion synthesis techniques based on motion capture data. In order to reuse mocap data in new scene, this paper deduced and realized the root locus edition method of virtual human. On the disposal stage at the end of motion synthesis, this paper proposed a posture-aligned smooth transition processing method between motion primitives. This method can reduce the abrupt change at the junction of two nerghboring motion primitives, thus making the movement transition of virtual human smoother and looks more natural.