Research On Human Motion Synthesis Driven By Local Features Of Data

Human motion synthesis is an important part in the area of computer graphics, which can be divided to model-based and data-driven ones. Compared to model-based ones, data-driven methods of human motion synthesis are more suitable for producing nature-looking whole-body human motions currently. Though, there are defects in mainstream data-driven technologies including too much manual customizations and high sample dependence. Considering local features of motions, some key issues in data-driven motion synthesis were explored and researched including motion transition generation, foot-skate free locomotion production and real-time control for virtual characters. In this thesis, we proposed Laplacian coordinates based motion transition generation, contact-based motion graphs and real-time foot-skate free locomotion synthesis supporting interative control.Motion transition generation is a basic technology in data-driven human motion synthesis. Laplacian coordinates based motion transition generation was proposed to reduce the manual customization and sample dependence. At first, motions were mapped to a curve segment in a multidimentional vector space, while each vetex represented mult-channel data in a frame and the neighborhood of each vertex represented the temporal relationship of the frames. Then the Laplacian coordinates of each vertex can represent the temporal local features near the time it formed. By connecting two curve segments representing two motions, a motion transition between these two motions can be formed from the result curve segment. Experimental results show that when the numbers of frames used to produce transitions in two input motions are decided, our method can generate smooth motion transitions with local features of both input motions. Without the need of arranging weights for each frames and the process of learning to find rules for generating new motions, this method possessed less manual customization and lower sample dependence.Foot-skate is a common kind of error when motions are synthesized from samples. To produce foot-skate free locomotions directly avoiding searching and optimizing on lots of samples, contact-based motion graphs were proposed and applied to generate locomotions for virtual actors. At first, information of the feet contacting with ground was used to build contact-based motion graphs. The nodes and edges in this graph differed in the possibility of changing motion direction while keeping poses and feet unchanged. According to contact information in these elements, new foot-skate free motions along user-defined paths on ground can then be generated directly by adjusting motion sequences exacted from the graph. Experimental results show that with the help of contact-based motion graphs, our method can produce foot-skate free locomotions for virtual actors with little samples, even when flying motion clips are included. Thus, this method can reduce the sample dependence dramatically as foot-skate cleanup based on searching or optimizing on lots of samples are no longer needed.Real-time controllable foot-skate free whole-body motions are difficult to synthesiz by mainstream applications currently. To make virtual actors can perform foot-skate locomotions according to user input in real-time, the method of motion synthesis supporting interactive control for virtual actors was proposed. At first, online path adaption was implemented based on the localty of adapting unit in path adaptation for contact-based graphs. Unadapted paths at different time were treated as different inputs. When an input was processed, at most one step of adaptation would be performed. By changing paths unadapted dynamically, users can control locomotions of virtual actors in real-time.Experimental results show that our method can make foot-skate free interactively controllable locomotions for virtual actors automatically. The manual customization in our method was reduced because there was no need to define feed-back strategies for virtual actors in many cases.As shown in our researches, with the help of local features of motions, the automation of data-driven motion synthesis can be improved for the reduction of manual customization and sample dependence.