Fast Vehicle Trajectory Data Prediction Based On Recursive Neural Network

With the rapid development of era,GPS(Global Positioning System)positioning has almost been covered fully in various fields.The application of GPS on the vehicle system is quite mature.The prediction of vehicle trajectory information is of great significance for the control of traffic flow,the safety of vehicle travel,and locationbased application service.This thesis predicts the vehicle trajectory information recorded by GPS based on recursive neural network methods the main work is shown as follows:(1)Vehicle trajectory data prediction based on SVD-LSTMAlthough the gate structure of LSTM(Long Short Term Memory)can store vehicle trajectory data for a long time,record the current state and cumulative historical state of vehicle trajectory selectively,which help it process the trajectory data of such a long sequence better.The gate mechanism of LSTM model for long-term and short-term state memory produces larger parameters(the largest proportion of parameters is the weight of network layer),resulting in slower training of the model.In view of the above factors,this thesis proposes a vehicle trajectory data prediction method based on SVDLSTM(VTDP SVD-LSTM).The method decomposes the LSTM layer weight into two smaller matrices by the SVD.Compared with the ordinary single-layer LSTM structure,it constructs one more custom layer.Through this idea of multi-layer grouping,the dimension of input layer and middle layer weight is reduced,with reduced computation of LSTM,and accelerated training speed of trajectory data.Experiments show that the training speed of SVD-LSTM is 2.32 times faster than LSTM in predicting vehicle trajectory data.At the same time,the accuracy of predicting vehicle trajectory points can be guaranteed as well,reaching 71.9%.(2)Vehicle trajectory data prediction based on FRU(Fourier Recurrent Units)Predicting vehicle trajectory data based on SVD-LSTM is a fast prediction from the external structure of recursive neural network.Although it can obtain better advantages in time,the accuracy of prediction is not high enough.Therefore,this thesis proposes a kind of recursive unit from the inside of the neural network,that is,changing the gate structure.Vehicle trajectory data prediction based on FRU can be quickly,while reducing the error and improving the accuracy.FRU is a recursive neural network with non-gate structure.Its Fourier residual learning structure can solve the gradient disappearance and explosion problems effectively in the training process,improving training stability and accelerating the convergence speed of the model.A summary of past hidden states of any vehicle can be approximated using a linear combination of Fourier statistics,resulting in a smaller error in the predicted trajectory points.In addition,due to the strong expression capability of the Fourier basis,fewer parameters can be used to express the vehicle trajectory data with less training time and predicting time.Experiments show that compared with the prediction results of other contrast models,the accuracy of FRU is the highest,reaching 76.8%,and the training speed of FRU is 3.27 times of LSTM.Therefore,the method can not only realize the rapid prediction of the vehicle trajectory data,but also deal with the long-term dependence of the vehicle trajectory data better,which preserves the hidden information of the vehicle trajectory completely.(3)Design and Implement a rapid vehicle trajectory data prediction prototype system based on recursive neural networkThis thesis designs a fast vehicle trajectory data prediction system based on recursive neural network,which can better show the prediction results obtained by the above two fast vehicle trajectory data prediction methods.The system can select the downloaded dataset stored in the database and provide the data preprocessing methods.After importing training set to the system,the model will be saved automatically after end of training.Then this system inputs the test set and loads the saved model to realize rapid prediction,which provides a good visualization for the presentation of the results.