Driving Style Recognition Method Based On Semi-supervised Learning Under Multiple Driving Conditions

Driving Style mainly refers to the driving habits of drivers,which is defined by the behavior characteristics of driving.Driving Style is closely related to intelligent transportation,driverless driving and insurance claims.Nowadays,there are many problems in the research of Driving Style recognition,such as the data source is not real,the factors are not comprehensive,and the driving style can not be grasped as a whole.In this paper,by collecting real driving data and taking the working condition as the minimum granularity of Driving Style recognition,a multi working condition Driving Style recognition model based on Semi-Supervised Learning is constructed.The main work is as follows:1.Build a database for Driving Style recognition.Through in-depth analysis of factors affecting daily Driving Style,a Driving Style acquisition test plan was formulated,and 80 test drivers were recruited to conduct two tests.The collected chassis CAN data,ADAS data,and driving environment data,etc.were divided into subjective data for comparison,the effective data segment is selected to build the Driving Style recognition database,which lays the foundation for the later construction of the working condition recognition system and the construction of the Driving Style recognition model.2.Established a complete working condition identification system.The entire driving process is divided into 4 longitudinal driving conditions and 4 lateral driving conditions.Through the analysis of the standard working condition data in the Driving style database,the physical quantity change curve template under each working condition is set and designed Working condition identification method based on logic and Dynamic Time Warping.After and experimental verification,the working condition identification system proposed in this paper has a precision rate of more than 90% under each working condition,and the obtained working condition results can be applied to actual problems.3.A multi-condition Driving Style recognition model based on the improved KL-Training algorithm is proposed.First,the Driving Style data is selected through linear discriminant analysis;secondly,the traditional Tri-Training algorithm is combined with relative entropy to solve the problem that the original algorithm may have a low degree of difference in the training set at the initial stage;finally,build different The semi-supervised learning driving style recognition model under working conditions,through the weighted decision fusion method,the Driving Style of each driver under each working condition is determined and integrated with the weight of the model accuracy,and the overall Driving Style of the driver is given.The experimental results show that the obtained Driving Style results reach a good accuracy rate,and they have the conditions to be loaded into the server for using.