Dynamic Estimation Of Queue Service Time At Signalized Intersections Using Lowpenetration Rate Vehicle Trajectory Data

Queuing information is the input information and evaluation index of signal control,which can reflect the operating state of an intersection.With the emergence of online-hailing car and connected vehicle,the form and content of vehicle trajectory data are increasingly rich.Estimating queue state based on vehicle trajectory data has become a research hotspot.Most of the existing researches focus on the aggregation process,using traffic flow theory or probability statistical methods to estimate queue length or the number of queuing vehicles.On the one hand,most of these methods have the assumption of arriving mode of vehicels,which may be inconsistent with the actual traffic flow.On the other hand,due to the lack of observable vehicle samples,most methods have poor performance in low-penetration rate environment.Therefore,from the view of vehicle dissipation,this paper proposes a method to describe queuing state by queuing service time,which means the time used to dissipate queuing vehicles in the green time.The method doesn’t need to assumpe arriving mode of traffic flow,and gives an estimation method under the condition of low-penetration rate.The main work of this paper includes vehicle trajectory feature information extraction,dynamic estimating model of queuing service time and method performance evaluation.In vehicle trajectory feature information extraction,according to the process of vehicle dissipation,the feature information mainly includes the vehicle departure time and queue state.In this paper,interpolation method and rule heuristic method are used to extract the two kinds of features.Aiming at the oversaturated scene,the free flow velocity of vehicles is estimated by Gaussian mixture model,and a method of extracting oversaturated vehicle samples is proposed based on the consistency of vehicle arrival and departure periods.The method is proposed to describe the queue service time needed to clear the stranded vehicles from the previous period.In the dynamic estimating model of queuing service time,in order to find the dividing point of queued and non-queued vehicles in the green time,this paper uses logistic regression model to describe the probability relationship between green light duration and vehicle queuing state.Considering the problem of insufficient samples of vehicle trajectories under the condition of low-penetration rate environment,this paper considers the correlation of queuing states at intersections in continuous periods,and adopts Laplace approximation method to construct a dynamic estimation method of queuing service time based on Bayesian theory.The posterior information of the estimation result of the previous period is regarded as the prior information of the next period,so as to realize the dynamic rolling of information.The cyclelevel dynamic estimation of queue service time is realized.In particular,when it is on oversaturated state,the queue service time will be more than the green time of the current cycle.This paper further extends the queue service time estimation model to the oversaturated scenario by considering the stranded vehicles from previous cycle.In the method performance evaluation process,this paper uses NGSIM vehicle trajectory data as the research data,designing evaluation tests with different penetration rate and different trajectory update interval.Meanwhile,we introduce two methods: maximum likelihood method without considering time continuity and direct observation method for comparison.The evaluation results show that the MAPE of the proposed method is 19.8% and 25.5%respectively when the vehicle trajectory penetration is 10% and 6%,the maximum likelihood method is 29.3% and 48.3% respectively,and the direct observation method is 40.0% and 58.3%respectively.Compared with the two comparison methods,the proposed method has better estimation accuracy under the condition of low-penetration rate environment.The results show that the estimation error of the proposed method increases about 4.2% when the update interval increases from 1 s to 5 s.When the update interval exceeds 10 s,the estimation accuracy of the proposed method is significantly affected.