Research On Data-driven Based Methods For 3D Human Skeleton Motion Synthesis

Human motion synthesis is one of the most important research fields in the field of virtual reality,which is widely used in the fields of entertainment,education and military affairs.The motion synthesis can be roughly divided into two categories: model driven and data driven,and the human motion synthesis based on the latter makes it easier to construct a high fidelity and natural human animation.At present,because of the high cost of motion capture technology,and the complexity of human motion,it brings a great obstacle to the reuse of motion data.In view of this situation,the paper focuses on the research of the key technologies and methods for the synthesis of human skeletal motion around the data driven human motion capture data.The main research content is as follows:1 ? Single motion sequence editing.Because each motion sequence's initial orientation,and translation are not the same.It is the pre-condition of human skeletal motion synthesis that to align the initial orientations of candidates motion sequences toward roughly the same direction.In this paper,we use the matrix rotation,translation and scaling operations to calculate the single motion sequence data,so that we can change the motion information of a sequence of motion in the virtual scene.2?Human motion similarity calculation.Only the motion sequences that have the potential for transition can be synthesized effectively.In this paper,the Euclidean distance is used to calculate the similarity of the motion sequences of same number of frames.The dynamic time warping algorithm(DTW)is used to calculate the similarity between the sequences of different number of frames.3?Experimental comparison of commonly used motion transition methods.Experiments are carried out based on the spatial position linear synthesis,the quaternion interpolation based on human motion posture and the interpolation algorithm based on inverse kinematics.The experimental results of the algorithms are compared and analyzed.We come to a conclusion that these methods are human-computer interaction intensive and sample data dependent.4?Motion transition editing based on multidimensional Laplace coordinates.3D Laplace coordinates can effectively express the local features of geometric data in 3D space,andthe dimension of human motion data is much higher.In this paper,we use the multidimensional coordinate to maintain the characteristics of the motion in time.Two curves of the motion sequences for synthesis are constructed,usingthe neighborhood to represent the temporal relationship between data,Multidimensional coordinate mapping method is applied to editand generate motion transition sequences.The experimental results show that the method of multidimensional Laplace coordinate motion transition is less human-computer interaction intensive,and the sample data dependence is also less.