Modeling And Analysis Of Car-following Behavior Using Data-driven Methods

The car-following behavior is an important phenomenon in road traffic. Especially in traffic congestion, the car-following behavior appears to be more common as overtaking maneuvers are relatively limited. The study on car-following behavior has always been a hot topic in the traffic flow theory. The car-following model is a useful tool to analyze the interaction of vehicles. By using the car-following model, researchers are able to understand the traffic flow characteristics and the spatial-temporal patterns of traffic congestion.Nowadays, the theory-driven method is widely used as the primary modelling method in car-following model. The theory-driven car-following model focuses on interpreting the influence factors of driving behavior. The modelers usually propose the assumptions of car-following model based on the daily observations. However, some of the researchers point out that the theory-driven car-following models might not precisely predict the actual car-following behavior. The prediction error shown in the error test is about 15% to 25%. Considering the complexity of car-following behavior, the theory-driven car-following model in fact cannot well predict car-following behavior by only using simple assumptions.With the development of intelligent transportation systems, one can obtain massive amounts of high-resolution vehicle trajectory data by using data collection techniques such as aerial photography, global positioning system and internet of vehicles. So it makes possible to use data-driven methods to acquire the valuable information inside of the vehicle trajectory data and build up car-following model with high prediction accuracy. This paper focuses on the modeling and analysis of car-following behavior using data-driven methods. The main research works are as follows:1. Modelling of car-following behavior based on data-driven methodsThe essence of the car-following model is to conduct the regression problem.Theoretically the data-driven model can provide excellent performance when solving regression problem. By reviewing and comparing the family of data-driven methods,this paper chooses support vector regression (SVR) as the main data-driven method for its higher prediction accuracy and establish a SVR-based car-following model.Moreover, by introducing the driving constraints, this paper proposes a new SVR-based car-following model which considers the driving constraints.2. Calibration and validation of the SVR-based car-following modelThis paper uses the field vehicle trajectory data provided by the Next Generation Simulation (NGSIM) project to calibrate and validate the SVR-based car-following model. It is proven that the calibrated model can well reproduce the traffic oscillation phenomenon, of which the occurrence and the spatial-temporal patterns can also be explained by performing the further simulation test.3. Analysis of the multi-anticipative driving behaviorSince drivers may anticipate traffic conditions farther downstream by considering the second leading vehicle and even more vehicles ahead in real world, it is necessary to perform an empirical analysis on the multi-anticipative driving behavior. Based on such an assumption, this paper extends the original SVR-based car-following to incorporate multi-anticipation parameters. In addition, the extended model can be used to perform a relative importance analysis that quantify the multi-anticipation in terms of the different stimuli to which drivers react in platoon car following.4. Analysis of the heterogeneity caused by vehicle typeSince real-world traffic flow usually contains a mixture of passenger cars and heavy trucks, it is necessary to perform an empirical analysis on the heterogeneity in car-following behavior caused by vehicle type. By incorporating the vehicle-type dependent parameters into the original SVR-based car-following model, a new model considering different vehicle types is proposed in this paper. The desired gap and the desired speed can be determined for each vehicle-type combination by using the SVR-based car-following model. The analysis result confirms the significant differences in desired gap and desired speed among different vehicle-type combinations. By performing a simulation test, the origin of the traffic congestion occurred in the mixed traffic flow can be well explained.5. Incorporation of the data-driven model with the theory-driven model The insufficiency in prediction accuracy of the SVR-based car-following model might be caused by the lack of training data. This paper proposed an extended SVR-based car-following model by incorporating the original SVR-based car-following model with the theory-driven model. The error test shows that the proposed model can provide higher stability of prediction accuracy than the original ones.6. The design and implementation of the vehicle trajectory analysis system This paper designs a universal data structure which is able to read arrangements of vehicle trajectory data from different sources. Based on that, a vehicle trajectory analysis system is presented and implemented by the functionalities such as the calibration of car-following model and the performance evaluation of traffic flow. This vehicle trajectory analysis system is intended to serve as a research platform for the analysis of vehicle trajectory data and data-driven car-following models.This study on data-driven car-following model would improve and supplement the current car-following model research level, and hopes to contribute to the development of intelligent transportation systems. In addition, the use of data-driven car-following model for traffic simulation can reveal the cause of traffic congestion and obtain the appropriate solutions that has a practical significance on reducing and eliminating traffic congestion.