Merging decision behavior model based on multivariate adaptive regression splines
[Objective]Weaving areas are bottlenecks of freeways,and lane-changing behavior is one of the main reasons for the capacity decline and traffic congestion in weaving areas.Frequent merging behaviors may lead to traffic flow disturbance upstream from the weaving area,affect the normal running of surrounding vehicles,and in severe cases may even lead to multi-vehicle accidents.An in-depth understanding of merging decision behavior in the weaving area is essential to reduce the vehicle collision risk and improve the traffic safety level.A newly developed nonparametric regression model—multiple adaptive regression splines(MARS)—is adopted to model the gap selection decision during merging to study the merging behavior in the freeway weaving area.[Methods]This study investigates complex interactions between merging and surrounding vehicles during merging.Trajectory data are extracted from the US-101 dataset provided by the dataset of next generation simulation program,and the symmetric exponential moving average filter method is used to smooth the data.Merging vehicles are influenced by surrounding vehicles in the auxiliary and adjacent main lanes.Thus,explanatory variables such as speeds,speed differences,gaps,and locations are calculated.Longitudinal and lateral collision risk indicators and time-to-collision are also considered to study the influence of collision risk on merging behaviors.Finally,925 observations are obtained and randomly divided into two subdatasets to train and test the model.The MARS model is compared with four state-of-the-art machine learning techniques:classification and regression tree,gradient boosting decision tree(GBDT),random forest,and logistic regression models.[Results]The speed difference between the merging vehicle and vehicles in the adjacent main lane played the most important role in gap selection.Interactions of influencing variables were observed.In particular,the best interaction level was 4 in the final model.The comparison showed that GBDT and MARS had the lowest rates of prediction error at 0.138 and 0.141,respectively.However,MARS could provide explicit expression functions that reflect the interaction between the influencing variables,which was beneficial to engineering applications.[Conclusions]By using the optimal variable transformation and potential variable interaction in the regression modeling scheme,MARS could easily handle complex nonlinear relationships in merging behaviors.This model could accurately predict the gap selection behavior and provide explicit expression functions,thus simplifying its understanding and application to driver assistance systems and autonomous driving systems.
NETL
NSTL
万方数据
维普
