Estimating link flows from limited traffic volume and sparse trajectory data: generative modelling approaches

Abstract

The accurate estimation of traffic flows (or volumes) on road links is critical in managing a roadnetwork and evaluating its performance. While loop detectors are installed to collect link flow data,the observation points are limited to a subset of links and there are still a large proportion of links thatdo not have direct observations. Thus, unobserved link flows need to be estimated based on availabledata and this is referred to as link flow estimation problem.This study addresses the link flow estimation problem by combining limited traffic volume dataand vehicle trajectory data. Vehicle trajectories detected by vehicle detection technologies (e.g., GPS,Bluetooth, video images, etc.) provide a promising additional data source, as they usually have betterspatial coverage than the loop detector data and offer deeper insights into vehicle propagationinformation. However, the observed trajectory data are usually sparse in that a trajectory sample onlyrepresents a small subset of the whole population. It is highly possible that the observed trajectorydata are not representative of the true population trajectory distribution. The main objective of thisstudy is to leverage these limited data sources (i.e., traffic volume data limited to a subset of links aswell as sparse vehicle trajectory data) to enhance link flow estimation in the road network.A novel generative modelling framework is proposed based on a Markov Decision Process(MDP), where an agent is created to make sequential decisions in the road network. The state-actionpaths sampled from this agent’s behaviour can be seen as synthetic vehicle trajectories in the roadnetwork. Reinforcement Learning (RL) methods are used to find the best agent’s behaviour, fromwhich the sampled synthetic vehicle trajectories are consistent with both the observed traffic volumesand the observed vehicle trajectories. Such synthetic trajectories describe a possible real-life trafficflow scenario that produces the observed traffic data. Therefore, link flow estimation results can beobtained from this scenario. To this end, two different RL methods are proposed and compared,inspired by the Inverse Reinforcement Learning (IRL) studies and the Constrained ReinforcementLearning (CRL) studies, respectively.The proposed generative modelling framework with either of these RL methods is validated bysolving the link flow estimation problem in a real road network. Additionally, experiment results onthe Nguyen-Dupuis network show the proposed methods can produce more reasonable link flowestimation results compared to methods in the literature, especially in scenarios where the observedvolume data are limited and the observed trajectory data are sparse.