Driver Lane-changing Behavior Prediction Based On Deep Learning

Vehicle lane changing is one of the most common behaviors during driving.Incorrect or untimely lane change is likely to cause traffic accidents or even casualties.If the driver’s lane change intention can be predicted in advance and the vehicle-to-vehicle lane change interaction notification can be performed using the Advanced Driver Assistance System(ADAS),the probability of accidents during driving will be greatly reduced.Against this background,this paper proposes a driver lane change prediction model for the purpose of improving prediction accuracy and perspective time.The main work of the paper is as follows:1)According to the current research conditions,a software and hardware platform capable of collecting vehicle kinematics data,driver kinematics data,and driver physiological data was developed and different drivers were invited to collect data.After the data collection is completed,the data is preprocessed.And a dynamic time window algorithm is proposed,which uses this algorithm to ensure the consistency of the model input data and improve the prediction time of the entire model.2)A time series prediction network based on a recurrent neural network(RNN)variant Seq2 Seq framework and LSTM and GRU calculation kernels was constructed.Because Seq2 Seq includes two parts,Encoder and Decoder,when constructing the network,four different network structures are LSTM-Seq2 Seq,GRU-Seq2 Seq,GRU-Encoder + LSTMDecoder Seq2 Seq,LSTM-Encoder + GRU-Decoder Seq2 Seq.A comparative analysis of training loss,fitting conditions,and convergence steps under the case of two-time windows shrinking is made.Through comprehensive consideration,it is found that the LSTM-Encoder + GRU-Decoder Seq2 Seq structure is most suitable as the basic network for time series prediction.3)The prediction results of Seq2 Seq are used to fuse and classify the data.In this chapter,three algorithms of deep neural network(NDD),support vector machine with Gaussian kernel function(GKF-SVM)and K-nearest neighbor(KNN)are selected to classify the data.By comparing the results of the three algorithms,it is found that the DNN results are performe well and can be used as a classification network.In addition,LSTM-Encoder + GRU-Decoder Seq2 Seq and DNN were cascaded,and the model was verified with new data.4)In order to find the input variables most relevant to lane changes,a sensitivity analysis was performed on the entire model.In sensitivity analysis,first add noise to the analysis variable or set it to 0,then observe the classification accuracy of the entire model,and use the average classification accuracy of the change as the sensitivity coefficient of the input variable.After calculation and analysis,it was found that steering wheel steering angle,β wave,steering angular velocity,and number of head rotations were the most relevant variables to the model.Finally,The Thesis concludes and looks forward.To sum up,based on real traffic scene data and innovative dynamic time window algorithm,this paper uses RNN-Seq2 Seq and DNN to construct a lane change prediction model,which achieves a prediction accuracy of 93.7% and an average perspective time of 2.2s.