Deep Learning-based Vehicle Lane Change Decision Recognition And Trajectory Prediction

The lane-changing behavior of vehicles is one of the main causes of traffic congestion and traffic accidents.The lane-changing behavior involves two stages,lanechanging decision and lane-changing execution,which is a complex process influenced by multiple factors.Traditional lane-changing models struggle to describe the effects of potential uncertain factors such as driver perception and driving style.Therefore,based on empirical data,this paper constructs a deep learning architecture for identifying lanechanging decisions and predicting lane-changing trajectories.This architecture can deeply extract the potential features of empirical data and solve the existing studies’ problems of low recognition accuracy and unsuitable long-term trajectory prediction,thus describing the real lane change behavior more accurately.The main research content of the paper is as follows:(1)Analysis of the lane change behaviour and extraction of relevant data to provide a foundation and support for subsequent model design and training.First,the paper divides the lane-changing process into lane change decision phase and lane change execution phase,and describe the lane-changing behaviour from the direction and speed aspects.Then,the influence factors of the lane change behaviour are analysed and determined.Finally,the NGSIM real traffic dataset is filtered,and lane change related data is extracted and labelled.(2)For the vehicle lane change decision recognition problem,this paper proposed a Conv2D-GRU based spatio-temporal attention(CGSTA)model.Different from the previous studies,this paper converts the lane change decision task based on numerical data processing into a video classification task for processing.The 2D convolution(Conv2D)and gated recurrent unit(GRU)are jointly used to deeply explore the potential spatio-temporal relationships of the data.Meanwhile,the spatial and temporal attention mechanisms are introduced in Conv2 D and GRU to enhance the model feature extraction ability,respectively.In addition,joint learning methods are used to mutually enhance the performance of the two decision tasks.The evaluation results of NGSIM data show that the CGSTA model is more excellent than Support Vector Machine(SVM),Multilayer Perceptron(MLP),Long Short-Term Memory(LSTM)and GRU in recognition accuracy,macro F1 score and prediction performance for lane change direction and speed decisions.The CGSTA model has excellent anti-interference and generalization abilities,and the individual component modules are effective for improving the recognition performance.(3)For the vehicle lane change trajectory prediction problem,this paper proposed a combined CGSTA and WaveNet-GRU(CGSTA-WG)model.Compared with the GRU encoder-decoder,the CGSTA-WG model employs CGSTA and WaveNet to capture the features of different types data respectively,and fuse the lane change decision recognition results.This combination can provide GRU with richer information and achieve more accurate long-term prediction.The NGSIM data evaluation results indicate that the CGSTA-WG model has a distinct advantage in long-term trajectory prediction compared to the LSTM encoder-decoder,1DCNN-LSTM,and GRU encoder-decoder.The model can achieve more accurate predictions with less input information and can better fit different types the real trajectory.The ablation experimental results demonstrate that the combination of CGSTA and WaveNet can effectively improve the prediction accuracy of GRU.The experiments further validate the CGSTA-WG model’s good anti-interference and generalization ability.