Extraction Of Vehicles’ Trip Chain And Estimation Of Traffic State Based On Quasi Population Trajectories

A comprehensive perception of the traffic operation status of the motor vehicle road network is a key prerequisite for solving the problem of urban traffic congestion and improving the efficiency of network operation.However,existing methods for estimating traffic states have problems such as sampling error,small spatial coverage,low resolution.With the development of artificial intelligence technology and information technology,the Automatic Vehicle Identification(AVI)system has been widely used in China.The AVI system not only realizes the functions of illegal capture,parking fees,etc.,but also provides valuable quasi-full-sample motor vehicle trajectory data in a city.This provides an opportunity to make up for the shortcomings in the existing traffic state estimation methods,and to further improve the traffic state estimation accuracy and quantitatively analyze the causes of traffic congestion.This research aims to extract the complete travel chain information of all drivers from the quasi-full-sample motor vehicle trajectory data recorded by the AVI system.It can estimate traffic flow parameters based on multiple measures including speed,flow,arrival(departure)time distribution,activity duration distribution,etc.Then both static and dynamic traffic status of the whole road network can be estimated.In order to achieve the above goals,this article has carried out research in the following aspects:(1)Modeling travel behavior for large-scale samplesExisting transportation choice behavior uses questionnaire survey as the main sample acquisition method,which has problems such as small sample size,high cost,and lack of information,which cannot support efficient modeling of large-scale sample transportation choice behavior.In response to this problem,by collecting AVI data with quasi-full sample characteristics and mobile communication network data,a modeling method for destination choice and route choice behavior for large-scale samples is proposed respectively.By considering the unique time and space characteristics of destination choice behavior,a destination choice model based on mobile communication network data is established,and an algorithm to estimate the constant parameter is given.Empirical analysis proves that the destination choice model can improve the rationality of the interpretation of behavior preference parameters and the model fitting.By integrating a small amount of labeled GNSS sample data and a large amount of unlabeled AVI sample data,a semi-supervised learning paradigm for route choice model is proposed.A comparative analysis is carried out based on the field-test data,which proves that the parameter estimation results of semi-supervised method can get the same reasonable behavior explanation conclusions as unsupervised and supervised methods.In addition,the semi-supervised method can significantly improve estimation accuracy.In terms of computational performance,the experimental results show that the proposed method has the ability to simulate large-scale driver route choice behavior.(2)Path recognition algorithm for sparse trajectoriesAlmost all existing path recognition algorithms(also known as map matching algorithms)are developed for high-resolution GNSS data and cannot be applied to sparse AVI data.Due to the large space and time interval,there are a large number of possible paths connecting AVI observation point pairs.Due to these uncertainties,it is very challenging to correctly identify the true path that matches the sparse AVI trajectory.AVI-MM,a path recognition algorithm customized for sparse AVI data,is thus proposed.This method first decomposes the AVI trajectory into consecutive pairs of observation points,and treats the travel between each pair of observations as an independent sub-trip.In addition to the conventional spatial analysis,dynamic time constraints and route choice models are also introduced to define the matching probability of sparse AVI trajectories.In addition,a data-driven candidate path set generation algorithm is proposed to generate a sufficiently realistic and attractive candidate set.The results of empirical experiments prove the three advantages of the proposed method:(1)Compared with the classic HMM algorithm,CRF algorithm and ST-MM algorithm,the AVI-MM algorithm significantly improves the matching accuracy without greatly reducing the computational efficiency.(2)The matching accuracy of 84.02%of testing samples exceeded 82.55%,which proves that the proposed method has better robustness to the input measured AVI trajectory.(3)The proposed candidate path set generation algorithm can achieve accurate replication of the observation path set.(3)Trip chain identification method incorporating both travel and activity behaviorThis paper proposes a method to identify the trip chain of vehicles through sparse and incomplete AVI trajectories.In order to effectively identify the activity locations,a problemspecific dynamic Bayesian network model named ST-II is constructed.Based on the proposed candidate mobility state generation algorithm and the optimal trip chain search algorithm,are developed to find the driver’s complete trip chain.The empirical results show that the identification accuracy of ST-II is much higher than that of traditional methods.In the sensitivity analysis of the size of the traffic area zone,it is found that there is an optimal value for it to achieve the highest identification accuracy.In addition,the experimental analysis also reports the analysis results of the impact of the spatial gap between the observation pairs.(4)Traffic state index system based on quasi-population trip chainsBased on the complete travel chain information of quasi-full samples of drivers,parking demand,traffic flow,and driving speed can be estimated through aggregate calculation.Then a series of traffic state indexes for static and dynamic traffic states are established.The experimental results show that the average absolute error of the estimated traffic flow on the entire network segment is 0.22.Further,the dynamic traffic state estimation results on the two scales of road link and township are presented.Moreover,the parking demand is also given based on the destination information in trip chains.With visual coding of the traffic status information,a web application is developed to motor the city-wide road network traffic state.Based on the collected real data sets,related algorithms and visualization interfaces were deployed and published online(access address is http://itse-seu.com/SY/daping.html).