Multisource Motion Retargeting For Human-like Characters

Human-like characters are the spirit of the three-dimensional (3D) animations. They are always in diverse styles and different structures. The design of their skeletal animations is facing two industrial problems, the high cost of motion data acquisition and the difficulty of directly reusing motion data, which restrict the efficiency of3D animation production. In order to overcome these difficulties and improve traditional processes of skeletal animation design, multisource motion retargeting method for human-like characters is proposed to explore a new way for more efficient production of various skeletal animations by existing motion data. In this dissertation, several key points are studied as follows:i) The concept and process of multisource motion retargetingThe traditional motion retargeting techniques transform motions from only one source character to the target character. They can neither maximize the reuse ration of motion data nor enhance the diversification of target motions. As a result, a novel method called multisource motion retargeting is presented. It transforms the motions of several source characters to the target character. So the target character can be rapidly driven by various motions. Accordingly, the efficiency of3D animation design is improved.ii) Automatic footprint detection based on spectral clusteringDue to the ever-present noise, it is difficult to directly detect footprints from motion capture data. A part of current footprint detection methods need interaction by user, and the other part are hard to be applied on mixed motion database. Therefore, we propose a novel footprint detection method based on spectral clustering. Firstly, the motion features of feet are represented as cluster samples. Secondly, the scaling parameter for each motion is selected by a self-tuning scaling selection method presented by ourselves. Finally, the motion frames are divided into the footprint frames and the non-footprint frames through spectral clustering.iii) Automatic motion retargeting between heterogeneous human-like charactersThe skeleton mapping of two heterogeneous characters is a basic problem in motion retargeting. It is difficult to achieve reasonable mapping results between two rather different human-like characters by current skeleton mapping methods. In view of this situation, we propose an automatic heterogeneous skeleton mapping method based on the hierarchical model of joints. The joints are layered in accordance with their importance. Besides, appropriate mapping strategies and mapping cost calculation methods are designed for joints of different layers. Then two heterogeneous skeletons construct their mapping relationship through the joints’ mapping layer by layer. Based on the mapping relationship, motion retargeting between heterogeneous human-like characters are achieved.iv) Diversified motion synthesis based on evolutionary strategiesAfter regarding motion clips as genes, a motion synthesis method is proposed by simulating biological evolution to generate diversified motions. In order to edit motion clips by different kinds of motion editing methods, three evolutionary strategies are defined, namely duplication, crossover, and mutation. In addition, a transition probability model of edited motion clips is built based on the probabilities of three evolutionary strategies. So that diversified motions with multiple source motions’ features can be synthesised by the edited motion clips.