On Vehicle Trajectory Prediction Model Based On Combined Clustering And LSTM Algorithm

In recent years,automatic driving technology has developed tremendously,automatic driving can reduce road traffic safety accidents caused by human factors and improving road traffic efficiency at the same time.However,there will be a highly uncertain and dynamic interactive complex driving environment with human and unmanned vehicles when an automatic driving vehicle enters the actual traffic scene.Automatic vehicles need to detect and collect the movement information of surrounding vehicles to ensure the safety and efficiency of vehicle driving,which requires real-time prediction of driving behavior and future trajectory of surrounding vehicles.It is a challenge to predict the surrounding vehicles future trajectory because of the limitation of surrounding vehicles information and the randomness of other vehicle movements.Based on the long short-term memory(LSTM)neural network model,the main research work in this thesis is as follows:This thesis constructs a CNN_LSTM(C_LSTM)vehicle trajectory prediction model based on convolution neural network(CNN)and LSTM.The prediction model consists of input layer,convolution layer,LSTM layer and full connection layer.The convolution layer network extracts the trajectory characteristics for the first time,which accelerates the training of the network model.LSTM network layer extracts the time series characteristics of the trajectory,and then expands the full connection layer to make the trajectory prediction.NGSIM highway datasets are used for the trajectory data,and the results shows that the average mean square error of the proposed C_LSTM model is better than the single-layer LSTM model.The C_LSTM model takes long time to train,and the interactive characteristics of vehicle trajectory cannot be expressed explicitly.In order to further improve the accuracy of the model and shorten the training time of the model,a clustering algorithm is proposed to mine the features of the trajectory data.The trajectory data features are analyzed by density clustering to get historical trajectory correlation,and the input trajectory segment is analyzed by fuzzy clustering to get the characteristic about driving behavior similarity.Interactivity information and different labels’ lane-changing information are extracted by clustering as new features of the data.By choosing the optimal input history trajectory length,input data is constructed with different feature combinations with interactive information.The model VC_LSTM,CC_LSTM,SC_LSTM with different combinations of features data are trainedIndependently,and those models are evaluated by MSE,RMSE,MAPE,DISTANCE,TIME indexes.Prediction models with interactive information and different combinations of features are evaluated and validated in both accuracy and time dimensions.The result shows that the SC_LSTM prediction model with surrounding vehicle interaction information and selected feature can meet the real-time and accuracy requirements of prediction.