Simulation And Generation Of Data-driven Virtual Human Motion

In the field of computer graphics,virtual human motion generation and simulation is an important research direction.Virtual human technology can be applied to game development,film special effects production and surgical simulation,which has very important application and research value.However,in these applications,the virtual person not only needs to accurately complete the action,but also needs to interact with the complex environment in real time,at the same time,to ensure the efficiency and authenticity of the virtual human action generation.In this context,this paper combines with the dynamic model or the deep learning model respectively,and attempts to realize the virtual human motion generation and simulation efficiently and realistically.The main research work and innovations of this paper are as follows:1.The existing virtual human dynamics controller based on classical control theory,the generated motion is mechanized and its realism is not high.Direct use of motion capture data for virtual human motion can not guarantee real-time interaction with the virtual environment.In view of the above problems,this paper proposes an algorithm framework based on kinetic energy model to analyze motion capture data.Firstly,the data is preprocessed,and the physical properties of the human body and the motion capture data are jointly calculated to obtain the control parameters with physical correctness;Next,the motion parameters are obtained by the motion splitter and the pose generator;Finally,the obtained motion parameters are obtained for controlling dynamic controllers.In order to verify the efficiency of this algorithm in the generation of virtual human motion and meet the interactivity requirements of complex virtual environment,real-time generation simulation of virtual human walking and running actions,and virtual people can realize conversion between walking and running.Further,the virtual human is placed in a complex terrain and it can interact continuously with the terrain in real time.Compared with other existing work,the proposed algorithm has advantages in data preprocessing,basic motion generation and complex environment interaction.2.For the complexity of virtual human motion information collection,this paper takes single-person video data as the data source and combines the deep learning related method to generate motion data,and proposes a complete algorithm from video data to virtual human motion data generation.Firstly,the stacking hourglass convolutional neural network is constructed to extract the heat map features of the image,and the two-dimensional spatial information of the human pose is obtained.Further,the integral regression method is used to estimate the three-dimensional spatial information.And further standardization design and optimization of three-dimensional spatial information is proposed.This paper demonstrates the results of different types of action generation to verify the effectiveness of this algorithm,and shows the estimation error compared with the existing methods.The error generated by the algorithm in each category of action has decreased.The algorithm of this paper has certain advantages over the existing methods.