Research On Quadruped Skeleton Movement Model Based On Mode Adaptive Neural Network

With the development of computer vision,skeletal motion model construction technology has a wide range of applications in special effects,game character control,human-computer interaction,virtual reality and other fields.The generation technology of human motion movements is currently developing very rapidly.But for quadrupeds,due to the complexity of their motions and the diversity of different subjects' movements,the construction of its skeletal motion models is still a challenging task.The existing prediction methods for motion are mainly based on search methods and based on real-time action generation.Among them,the search-based method requires the support of a huge animation database,and the most suitable next frame of animation is found from the database according to the control signal combined with the current trajectory,posture,and speed of the model.The method based on real-time action generation also requires a large amount of structured motion capture data.At present,most of the skeleton data collection uses depth cameras and depth sensors.It is often not so easy to obtain,especially the structured motion capture data of quadrupeds.In view of the problems existing in the construction of skeletal motion model,this paper realizes a data-driven motion prediction method based on the timing of motion data and the diversity and complexity of quadruped motion.The main research content is as follows:(1)Processing the original skeletal motion data.Extracting the required data from the bvh file with skeletal information,building skeleton model in Motion Builder based on the skeletal information in bvh,importing skeleton models and bvh files into Unity,completing the conversion of skeletal motion information from the bvh file to world coordinates,and then extracting the corresponding features to represent the current motion posture.(2)A prediction method based on LSTM long-short memory neural network.After analyzing the time series properties of the quadruped motion data,an algorithm based on the Long Short-Term Memory neural network model for motion prediction is proposed,which learns the trend of motion data in the real environment through the Long Short-Term Memory neural network,and predicts the motion data of the next frame and compares it with the real data.Then analyze the network structure of the MANN(Mode Adaptive Neural Network),adjust its network structure and optimize the parameters.Compare the experimental results of the LSTM prediction model,the MANN prediction model and the adjusted MANN prediction model.(3)A hybrid model combining LSTM(lonug and short-term memory)neural network integrated into the attention mechanism and MANN(mode adaptive neural network).After analyzing the principle and structure of the LSTM prediction model and the MANN prediction model,using the idea of dynamic network weights in the mode adaptive neural network,based on the LSTM neural network,and the weight of the neural network is dynamically given by the Mixture of expert.And through experiments to prove the superiority of the LSTM neural network integrated into the attention mechanism in motion prediction.Finally,a skeletal motion prediction model based on the LSTM neural network,introduced the attention mechanism and Dropout,combined with Mixture of experts,is established.And through experiments,it is proved that the hybrid model can improve the prediction performance of quadruped.