Research On The Construction Method Of Personalized Driver Model Based On Driving Style Transfer Network

Driver model is one of the basic research contents and key links of automobile traffic safety, intelligent transportation system, unmanned driving and other technologies. With the rapid development of automobile intelligence, automatic driving technology with different levels and different scenes will be gradually applied in practice, and people’s demand for personalized driver model which can reflect driving habits becomes more and more urgent.However, the related research on this personalized driver model is still lacking at present, and the occurrence of traffic accidents is often related to the driver’s personalized manipulation behavior, and a bad driving habit can also lead to traffic accidents on specific roads. In order to meet the above requirements, this paper proposes a personalized driver model framework based on driving style migration network. Under this framework, a part of historical data of other drivers is selected, combined with a small amount of target driver data as input, and a driving behavior data migration model conforming to the target driver is formed through the driving style migration network to extract the essential characteristics of the driver. With this model,a large number of migration learning data in line with the target driver can be obtained,so as to train the target driver model, complete the personalized driver behavior modeling on the data level, and realize the long-term prediction of the driver’s lane change behavior. The main research contents are as follows:(1)Construction method of personalized driver model based on driving style transfer network. Firstly, this paper introduces the ideological source of the personalized driver model construction method based on driving style migration network, then constructs the personalized driver model framework based on driving style migration network, and finally briefly describes the key technologies in the overall thinking of the above model construction method.(2)Establishment of data migration model based on driving style migration network. In view of the limitation of specific model structure in the modeling process of traditional personalized driver behavior modeling method, which can not make full use of historical data,this paper analyzes the advantages and disadvantages of existing data migration methods, and proposes a data migration learning model construction method based on driving style migration network. Through sufficient driving behavior data of other drivers, combined with a small amount of target driver’s operation information, vehicle state information, Transfer the driving behavior data of other drivers to the target driver, complete the knowledge transfer among different driver data, realize the reuse of historical driving behavior data, and provide sufficient data support for the subsequent personalized model training of the target driver.(3)Establishment of driving behavior prediction model based on LSTM-MLP. The prediction model includes driving behavior prediction network and test network. In view of the shortcomings of traditional neural network in time series data prediction, this paper first establishes a driving behavior prediction network according to the encoder-decoder structure,and builds a test network based on multi-layer perceptron; Then the joint loss function is constructed as the loss function of the driving behavior prediction model to meet the needs of driving behavior prediction; Finally, the prediction model is trained and verified by the preprocessed data of lane change behavior.(4)Experimental verification of personalized driver model based on driving style transfer network. Firstly, through analyzing the requirements of personalized driver model, an experimental platform of personalized driver model is designed and built, and then the platform is used to collect and preprocess the driving behavior data in the lane change scene.Finally, the training and prediction of personalized driver model are completed by using the real vehicle data, and the experimental results are analyzed to verify the effectiveness and accuracy of the proposed personalized driver model construction method.