Human Motion Synthesis Method Of Machine Learning Research

With the rise of digital media entertainment industry which represented by3D movies and games, computer graphic has developed extremely rapidly and it is of more and more particular interest. Among a number of motion generation technologies, human motion capture is most widely used in the field of science and industry. However, there are also several problems with human motion capture, such as the expensive capture cost and difficulty to reuse captured data. How to reuse the captured data and create new motions is great valuable for science and economy.Human motion synthesis is that we modify the existed motions in the motion databases to create new motions which met the requirement. In this thesis, we propose semantic parametric models of motion data by component analysis methods which are parts of machine learning algorithms. The high-dimensional motion squences are controlled by the semantic parameters of low-dimensional spaces. By modifying the low-dimensional parameters, realistic and natural motion sequences are intuitively synthesized in real-time. The work of this thesis mainly includes the following3aspects.1. The motion style transference method based on independent component analysis.Taking frames in single motion sequence as samples, we use independent component analysis to process the data combined by two similar motions which have different styles. Among all the independent components, the ones which changes with large differences from the others are selected and transferred between the two motions to synthesize new motion sequence possibly if they maintain proper styles. The transference of motion styles using the algorithm is comfirmed by the experiment.2. Human motion synthesis based on principal component analysis and block principal component analysis. The algorithm based on principal component maps the high-dimensional similar motions to low-dimensional spaces and synthesizes new motions by change the value of parameters of low-dimensional spaces. The problem of the algorithm is that it is impossible to understand the parameters of low-dimensional spaces. In this thesis, a motion synthesis method based on block principal component analysis is proposed. The motion data is divided into several groups according to the structure of human skeleton. After applying block principal component analysis to each group, the parameters of the low-dimensional subspaces have understandable semantics. By intuitively adjusting the semantic parameters, desired new motions could be synthesized in real-time. 3. Human motion synthesis based on independent component analysis and block independent component analysis. The procedure of motion synthesis based on independent component analysis is similar to the algorithm based on principal component analysis. After applying independent component analysis to motion data, the parameters of the low-dimensional subspaces have understandable semantics. We improved the algorithm and proposed a motion synthesis based on block independent component analysis by combining independent component analysis and grouped motion data in this thesis. The method raises the control of local motion details by increasing the number of independent components in some groups. The experiment result proved that the improved method meets the requirements and reduces the complexity of interaction at the same time.