Vehicle Trajectory Prediction Based On RNN-LSTM Network

In order to alleviate the current traffic jams,people began to explore automatic driving.An indispensable part of vehicle intelligent driving is the accurate prediction of vehicle trajectory.The prediction results affect the accuracy of vehicle route planning.There are two main directions for vehicle trajectory prediction: traditional method and neural network model.Although the traditional trajectory prediction method is earlier than neural network,the huge computing power of neural network and the success of pedestrian trajectory prediction make researchers try to apply neural network to vehicle trajectory prediction.On the basis of previous studies,this paper uses RNN and LSTM unit to build prediction model to improve the accuracy of vehicle trajectory prediction.The main contents of this paper are as follows:First,segment division and calibration of lane change.This paper uses symmetric exponential filtering algorithm to preprocess the original data and extract the effective data.Then,the intersection of the vehicle trajectory and the lane line is marked as the lane change point,and then the lane change point is taken as the base point to delimit2.5 seconds forward and backward respectively,forming a 5-second lane change segment.According to the calculated vehicle heading angle,the lane change type is judged by the positive and negative value of the heading angle.Secondly,the sample sequence of model training is extracted.The training sample sequence is calibrated by sliding window method.The length of sliding window is set to 40,and the training historical track sequence is extracted by sliding backward in 0.2second time step.Each sample sequence contains 40 historical information sampling points and 50 prediction information sampling points.Then,the RNN-LSTM network model is constructed.According to the basic LSTM network for vehicle trajectory prediction,the model only inputs the target vehicle’s own parameters,and the result is called R＿LSTM。 Adding the vehicle information of surrounding environment into the model,the training result is called B＿LSTM model results.On the basis of previous research,this paper builds RNNLSTM network trajectory prediction model.The analysis shows that RNN-LSTM has obvious advantages in long-term prediction accuracy and higher accuracy.Finally,according to the RNN-LSTM neural network model built in this paper,the influence of different prediction time on the accuracy of trajectory prediction is analyzed.Through the study of specific vehicles,it is found that the probability of correctly identifying lane change trajectory is less than 50% when the prediction time is more than 3 seconds.When the end point of the historical sequence segment is 1.5seconds before lane change point,the prediction accuracy of RNN-LSTM model can reach more than 80%;When the starting point of sequence segment is one second before lane change point,the prediction accuracy of the model has exceeded 83%;When the starting point of sequence segment reaches the lane change point,the prediction accuracy of the model is more than 87%.In addition,the RNN-LSTM network model built in this paper can predict the lane changing intention of vehicles on the line segment with a prediction time of 3 seconds and outside the lane changing segment trajectory.