In recent years,robot and its related technologies have been developed rapidly,especially the interactive tasks between robot and environment,such as capturing and intercepting free-flying moving objects,are widely used in aerospace,intelligent military,advanced intelligent manufacturing and sports event assistance and training,The prediction of the orientation and position of moving objects is the precondition for the robot to complete the above interactive tasks.In this thesis,based on five kinds of objects,the Long Short-Term Memory(LSTM)network model and the LSTM network model with spatio-temporal attention mechanism fused are established respectively to study the real-time orientation and position prediction of moving objects.First of all,this thesis uses the Vicon motion capture system to collect the trajectory data of five kinds of moving objects through human throwing demonstration,and the trajectory data are enhanced by the data preprocessing,thus the richness of the trajectory data is improved,so that the trajectory data set of moving objects with 5000 tracks is constructed,which lays a foundation for the research on the prediction of orientation and position based on neural networks.Then,in order to solve the problems that the kinematics model is difficult to establish,the model parameters are difficult to identify and the model is not generalized,this thesis establishes a moving object orientation and position prediction model based on LSTM network.By using a large number of LSTM units,the changing characteristics of orientation and position in the process of object motion can be addressed,and the timestep of the key parameters of the model is optimized,and the trajectory data set of the moving objects constructed in this thesis is used to train and test the model.The experimental results show that the model can achieve good prediction results and meet a certain real-time performance.In addition,since the trajectory information at different historical moments generally leads to different influences on the future trajectory change trend of moving objects.In this thesis,a spatio-temporal attention network model for orientation and position prediction is established,which fuses the spatio-temporal attention mechanism with LSTM network.This model studies the temporal and spatial dependence characteristics in the process of object motion,that is,an attention mechanism is introduced to analyze the influence of trajectory information at different historical moments on future trajectories,and the LSTM network is used to learn the characteristics of spatial orientation and position changes in the process of object motion.In order to verify the performance of the established network model,the average displacement error,the final point displacement error and the average time of predicting a complete trajectory are taken as the evaluation indexes,and the conventional linear model and Recurrent Neural Network(RNN)network are taken as the benchmark model,which are compared with the proposed LSTM model and the spatio-temporal attention network model.The experimental results show that the orientation and position prediction model of moving objects based on spatio-temporal attention network has higher prediction accuracy,which is better than other similar methods,the average time of predicting a complete trajectory is slightly increased in comparison to LSTM network.In addition,the test results of five different objects motion prediction show that the two proposed network models have a certain generalization,and can be used to model and predict other objects with time series motion characteristics. |