Trajectory Prediction Of Moving Target Based On Recurrent Neural Network

The capture and interception technology of moving targets has been widely used in military,aerospace,industrial production,sports and other fields.Recognition,location and trajectory prediction of moving targets are the basis and prerequisite for capturing and intercepting them.The traditional methods of recognition and location methods depend strongly on the environment;the traditional methods of trajectory prediction are mainly based on kinematics model and have poor generalization ability.Therefore,this thesis uses depth camera to acquire scene depth images,and adopts the background subtraction method based on Gauss Mixture Model to realizes real-time recognition and location of the moving targets,and adopts Recurrent Neural Network(RNN),which has good environmental adaptability and generalization ability and is suitable for time series prediction,to carry out relevant research on motion trajectory prediction.The main tasks completed are as follows:(1)System platform construction.The hardware platform of the system is built with Kinect depth camera and NOKOV motion capture system,and ROS(Robot Operating System)system is used as the software platform of the robot.The system is calibrated,including the internal parameter calibration of Kinect depth camera and the joint calibration of Kinect and NOKOV motion capture system.(2)Recognition and location of motion target.The background subtraction method based on Gauss Mixture Model is used to recognize moving targets;the phase plane's location of moving targets is used to obtain point cloud information of corresponding pixels;the centroid of point cloud is fitted by Gauss Newton method to obtain the spatial coordinates of moving targets;the motion trajectory of the center of mass was optimized with kalman filter,and the target location error is reduced from 13 mm to 10 mm.(3)Establishment of trajectory prediction network for moving target.The long Short Term Memory(LSTM)network model is built by using RNN-based trajectory prediction method for moving targets.The step length L,the number of hidden nodes H and the learning rate ? of the network model are optimized by using the model parameter optimization algorithm based on multi-layer grid search.Finally,Back Propagation through Time(BPTT)algorithm is used to optimize the weight parameters of the LSTM network model.(4)Experiments and results' analysis of trajectory prediction for moving targets.Firstly,the key parameters of the model are optimized,and the step size L,the number of hidden nodes H and the learning rate ? are determined to be 25,50 and 0.07 respectively.Then,the trajectory prediction experiments of moving objects are carried out.The trajectories of moving objects are predicted by standard RNN model,LSTM model and GRU model respectively,and the prediction accuracy of the three models is compared.When predicting the trajectory landing point of the moving target,the standard RNN model cannot make effective prediction,while the prediction accuracy of the LSTM model is 45.36% higher than that of the GRU model.Finally,the built network is applied to the trajectory prediction of irregular target table tennis bats,which proves the generalization ability of the built network.