Research On Lane Changing Intention Recognition And Lane Changing Trajectory Prediction Based On Machine Learning

It is well known that highway traffic accidents have posed a serious threat to the safety of drivers and passengers.Further studies have also shown that vehicle lane changing behavior is one of the main causes of traffic accidents.Each lane change behavior is not only restricted by the driving conditions of surrounding vehicles,but also affects the driving of surrounding vehicles.In this thesis,the NGSIM trajectory dataset was used as the research data and then the vehicle location and velocity distribution characteristics of the dataset were analyzed.Before the start of the research,the noise affecting the accuracy in the original dataset was processed by wavelet noise reduction method to lay the foundation for the training and verification of the later model.According to the model training requirements,the data was extracted and reconstructed,and then the Gaussian Mixture Model-Hidden Markov Model(GMM-HMM)was used to establish the lane changing intention recognition model.Considering the contribution of different features in the trajectory data to lane changing intention recognition,lateral velocity and lateral location were selected as observable states,and the model parameters that best describe the relationship between observable states and lane changing intention were estimated iteratively,then the recognized accuracy was verified and the causes of identification delay was analyzed.On the basis of lane changing intention recognition,a Gated Recurrent Unit(GRU)neural network with the sliding time window method was used to establish the left and right lane changing trajectory prediction model,and the optimal combination of hyper parameters was determined using RMSE.After that,the prediction results were verified and evaluated under the test dataset.Then,by comparing the prediction effects under different time windows,the differences and the advantages and disadvantages of using long time windows and short time windows for prediction were obtainedFinally,in order to reflect the actual application value of the lane changing intention recognition-lane changing trajectory prediction model,a case application analysis was carried out in the context of car lane changing collision risk recognition.After determining the direction and the starting point of lane changing by the lane changing intention recognition model,the lane changing trajectory of the target vehicle was predicted and the driving information of neighboring vehicles was extracted,through which the distance between target vehicle and neighboring vehicles could be calculated.Then the risk level was determined according to the Time Headway(THW)and the relationship on highway cross section between the neighboring vehicles and the target vehicle during the lane change process.The research results show that the GMM-HMM lane changing intention recognition model can accurately recognize the driver’s driving intention with both accuracy and precision are more than 90% when using trajectory data;the GRU neural network can predict future driving trajectory using historical trajectory data by learning the characteristics of lane changing,when using the 2s time window to predict the 6s lane changing trajectory in the future,the average prediction deviation distance at the 6th second of left and right changing trajectory are 6.05 m and 4.41 m respectively,and the training speed is better than Long Short-Term Memory(LSTM)neural network;the longer historical trajectory of the lane change used for prediction,the prediction accuracy is higher;in the application background of lane changing collision risk identification for car,the established models can not only accurately identify the collision risk level of the target vehicle at the starting point of the lane changing behavior,but can also identify and explain the risk changing during the lane changing process.