Efficient Traffic Estimation With Multi-Sourced Data by Parallel Coupled Hidden Markov Model

Abstract

Traffic congestion estimation in arterial networks with sparse GPS probe data is a practically important while substantially challenging research issue. The effectiveness and efficiency of the existing GPS probe data-based traffic estimation models are largely limited due to the following two challenges. First, due to the low sampling frequency of GPS probes, probe data are usually sparse, especially for some road links not located in the central urban areas. Second, due to the very complex temporal and spatial dependencies among the road links, the variable space of the existing traffic estimation models is huge. It is time consuming to get an accurate estimation of a large arterial road network with thousands of road links. To address the above-mentioned issues, this paper proposes to extract traffic event signals from social media and incorporate them with GPS probe data to alleviate the data sparse issue. We first collect traffic-related posts that report various traffic events, including traffic jam, accident, and road construction from Twitter. By considering the GPS probe readings and the traffic event tweets as two types of observations, we next extend the conventional coupled hidden Markov model for integrating the two types of data to obtain a more accurate estimation of traffic conditions. To address the computational challenge, a parallel importance sampling-based electromagnetic algorithm is further introduced. We evaluate our model on the arterial network of downtown Chicago. The experimental results demonstrate the superior performance of the model in both effectiveness and efficiency.