Research On Driving Style Recognition And Decision-making Of Intelligent Car Following Behavior

With the rapid growth of car ownership and the number of drivers,problems such as road congestion,environmental pollution,and road safety have become increasingly prominent.To improve the current traffic situation,advanced driver assistance systems are gradually promoted and applied to vehicles.For example,longitudinal driver assistance systems can warn the drivers when drivers have operating errors or take over the vehicle when conditions permit,which helps drivers maintain safe and comfort in car-following events.However,the existing longitudinal driving assistance systems take the same approach to help drivers,and rarely consider the differences in drivers’ driving styles,which leads to low acceptance,trust,and willingness of drivers for using it.If the longitudinal driving assistance system can recognize the driving style of drivers and make decisions according to their needs and preferences,it will help to improve the comfort of the vehicle,thereby increasing the driver’s trust and acceptance of the system.Therefore,carrying out driving style evaluation research on car-following events and analyzing different driving style preferences,then applying it to driving style recognition and car-following decision-making model have important significance both to theory study and engineering guidance.This paper evaluates the driving style of vehicles at car-following events,and on this basis,constructs a driving style recognition model and personalized car-following models.The main works are as follows:(1)The paper proposes a driving style evaluation method for car-following events.We extract stable car-following events from the natural driving data set as research samples,then select multi-feature parameters to represent the driving style,and reduce the dimension of parameters through the principal component analysis algorithm.After that,we build a driving style scoring model by using the first main component and then calculate the driving style score of samples.Finally,the Gaussian mixture model is used to cluster the samples based on the style scores to obtain the driving style of the samples.The results show that the Gaussian mixture model algorithm attains a good clustering effect.At the same time,compared with the method of evaluating the driving style directly based on the clustering of multiple principal components,the driving style clustering based on the scoring model achieves a more reasonable driving style evaluation results:The time headway and jerk of samples were significantly different with different driving style,which provided data support for follow-up research.(2)Driving style recognition models based on sample time-series features are constructed.Construct the time series feature set of car-following behavior based on the headway and jerk of the sample,and use this as input to construct a driving style recognition model based on convolutional neural networks(CNN).In order to take full use of temporal information of the feature set,the Gaussian process classifier(GPC)is used to replace the classification layer in the CNN models to construct a CNN-GPC model.Comparing the two driving style recognition models with models based on support vector machines and random forests,the results show that the CNN-GPC model can make better use of the time series information of the samples,which helps to attain better driving style recognition results than the other three models.Meanwhile,the CNN-GPC model has good classification performance.(3)Three car-following models with different driving styles are established.First,model the decision-making process of car-following behavior as a deep reinforcement learning problem,and use the proximal policy optimization algorithm to solve the problem.Based on these,the basic structure of the carfollowing model is designed.Then,according to the mean value of headway and jerk of different driving style samples,the model’s reward function parameters are designed,which are used for establishing car-following models with different driving styles:conservative,normal,and aggressive.The results of numerical simulation experiments show that:the car-following models with different driving styles show corresponding style characteristics to a certain extent in the process of car-following,and all can complete the car-following task relatively safely.