Study Of Car-following Behaviors Considering Unstructured Data And Driver Characteristics

Car-following model is an important part of microscopic traffic flow theory.It provides theoretical support for microscopic and macroscopic traffic flow simulation,and is also widely applied in the fields of capacity analysis,traffic safety evaluation,active safety control,automatic driving,and driver assistance system.In the context of abundant traffic big data resources and further development of traffic data analysis and mining technology,it is of great significance to reshape the modeling ideas and realize fine modeling of the car-following models.The existing studies are slightly lacking in the aspects of following behavior portrayal and model input selection,driver characteristics portrayal and fusion,evaluation system construction,and new requirements brought by the development of vehicle automation.Therefore,this study focuses on the study of following behavior considering unstructured data and driver characteristics,and constructs a data-driven following model based on a deep learning framework.This paper firstly studies and analyzes the main problems existing in the NGSIM trajectory dataset.Through Kalman filtering method,the data pre-processing is carried out on the basis of ensuring that the system conforms to its dynamic law.Data quality is improved.An analysis perspective and a reference method are provided for actual trajectory data processing.Then,the evaluation system is constructed.The two main tasks of the car-following models,i.e.,one-step prediction and multi-step prediction,are evaluated from multiple perspectives.In the evaluation of single-step prediction,this paper focuses on the accuracy and stability of speed and position prediction;in the evaluation of multi-step prediction,this paper designs four evaluation indexes with different focuses,which are absolute accuracy measure of position,reliability measure,accuracy measure insensitive to the length of spatio-temporal trajectory,and long-range prediction effect measure.In order to more comprehensively,intuitively and accurately portray the driver’s perception of the surrounding road traffic conditions,unstructured data is introduced into the car-following model and coded based on the information entropy theory.The car-following model considering unstructured data(DIMNN)performs well in one-step prediction stability and outperforms the baseline models in multi-step prediction tasks.Finally,considering the influence of driver characteristics on the following behavior,this study proposes a following behavior modeling method that incorporates driver response delay and driving style analysis.The car-following model considering driver characteristics(DGRU)performs well in the one-step prediction task and slightly outperforms the DIMNN model in the multi-step prediction task.The car-following model considering unstructured data and driver characteristics(DIMGRU)is a fusion of DIMNN and DGRU models.This model improves the lag of acceleration and deceleration behavior prediction,and the volatility of speed prediction is significantly reduced.The multidimensional mining of car-following behavior in this paper provides a new perspective and dimensionality for related research.Also,it provides solutions for fusing and processing different types of data for autonomous vehicles.It provides a path for the differentiated and refined analysis of different human drivers for autonomous vehicles or assisted driving systems.