Research On Human Skeleton Data Modeling And Motion Transition Via Deep Spatio-temporal Feature Mining

Since the advent of motion capture technology,it has had a profound impact on all aspects of animation production at the time and current film production,virtual reality,robot control,medical health and so on.With the advent of the era of artificial intelligence,self-driving cars and service-type intelligent robots need to have certain behavioral interaction capabilities.Human motion predictions can detect potential actions in advance to provide a seamless interactive experience.In animation creation and movie special effects production,in order to make full use of motion sequences,short-term,single-semantic sequence fragments are smoothly stitched into a long-term motion sequence with special meaning.In view of the above two types of problems,this article focuses on two aspects of deep learning,the establishment of a neural network prediction model and a bidirectional prediction fusion model.Thereare study achievementsas follows:(1)Attention-based neural network prediction model: Human skeleton movement data is a continuous sequence in the time domain,and LSTM is very suitable for processing time series data,the difference is that the attention layer is built before each recurrent neural unit.The motion frames of each time step are first assigned the attention weights of each joint point in a single frame via the attention layer,extracting obvious or implicit features in the motion sequence,and then the recurrent neural unit stacks the motion frames in the time domain feature.To improve model performance,add residual connections in the decoder section.The experimental results show that the network model has excellent prediction performance for human motion prediction tasks.(2)Neural network prediction model based on graph convolution: The movement of natural humans remains continuous in time and space,while the previous work ignored the spatial characteristics of human movement.Therefore,in this section,a graph convolution layer is established,and the 17 joint points of the human body are respectively calculated in the graph convolution layer for the spatial relationship with their neighbor nodes.When the recurrent neural unit advances one time step forward,the graph convolution operation is performed to achieve the purpose of simultaneously extracting the motion characteristics of the motion sequence in the time domain and the space domain,and improving the prediction performance of the model.(3)Bidirectional prediction fusion model of human motion sequence: Aiming at the task of splicing human motion sequence,a bidirectional prediction fusion model is proposed.Due to the performance of the sequence prediction model increases with the time goes by.The errorspredicted become worse and worse,so the framesof predicted are given different weight coefficients according to the time distance from the last frame of the ground truth sequence.In the sequence fusion stage,a weighted average fusion method was proposed,and the weighting coefficient of each predicted frame was used to smooth the sequence of motion transitions.