| Vehicle lane changing is a common behavior in the driving process.Improper lane-changing behaviors may lead to traffic accidents,which not only affect road operation efficiency but also threatens people’s lives and property safety.How to identify the lane-changing intention of drivers and forecast lane-changing trajectories accurately has become a hot issue in the field of traffic safety and intelligent transportation.Therefore,this paper build models based on vehicle trajectory data to identify the lane-changing intention of drivers and forecast the lane-changing trajectories.Firstly,the Ubiquitous Traffic Eye dataset is used as the research data in this paper.After analyzing the errors and noise in the original data,the LOWESS algorithm and s EMA algorithm are used to smooth and filter the longitudinal and lateral data respectively.On this basis,the vehicle heading angle is selected as the criteria for extracting suitable vehicle lane change trajectory data to provide training and validation data sets for subsequent studies.Next,based on the pre-processed lane change trajectory data sets,the drivers’ style features are extracted by analyzing the driving parameters of vehicles,and the drivers are classified into three categories: aggressive,general and cautious by using the Gaussian mixture model.The clustering results are used as the driving style label values of drivers,and the vehicle lane changing trajectory data are used to train and test the driving style recognition model of drivers.The results show that the support vector machine model based on the RBF kernel function has the best recognition results,and the accuracy rate is higher than 95% for all three types of driving styles.And then,the vehicle lane-changing trajectory data and lane-keeping vehicle trajectory data are reconstructed and spliced.The concept of vehicle interaction area is proposed,and the influence of surrounding vehicles on the target vehicle is attributed to the function of the overlapping area of vehicle interaction area.With the driving parameters of the target vehicle as the feature parameters,the XGBoost lane-changing intention recognition model is established.By comparing the recognition results of different lengths of lane-changing intention time windows,the results show that the prediction accuracy is higher than 95% when the time window length is less than 2.5s.Based on the appropriate time windows,the results show that the lane-changing intention recognition accuracy is improved by considering the driving style of drivers,and the overall recognition accuracy can still reach over 96% with a lane-changing intention time window of 2.5s.Finally,based on the results of lane-changing intention recognition,a hybrid XGBoostLSTM model of vehicle lane-changing trajectory forecast is established.The results of lanechanging intention recognition are output by the XGBoost model as new parameters input to the LSTM lane-changing trajectory forecast model,and together with the vehicle trajectories as historical data to forecast vehicle lane-changing trajectories for the future.The layer-optimal network structure model was found by constructing LSTM network models with five different layers.Taking the driving style of drivers into account,it was found that the prediction accuracy of the lane-changing trajectories could be improved,and the prediction of vehicle trajectory for the next 3s with a sliding time window length of 1.25 s improved the fit by 6.98%for rightward lane-changing behavior and by 6.21% for leftward lane-changing behavior.In addition,the predictive deviation of the lane-changing trajectory increases gradually with the prediction time due to the accumulation of errors.Through this paper,it is concluded that it is realistic and feasible to improve the accuracy of intention recognition and trajectory prediction by considering the relationship between the driving style of drivers,lane-changing intention and vehicle lane-changing trajectory comprehensively. |
