Urban Traffic State Estimation with Online Car-Hailing Data: A Dynamic Tensor-Based Bayesian Probabilistic Decomposition Approach

Abstract

Timely and precise traffic state estimation of urban roads is significant for urban traffic management and operation. However, most of the advanced studies focus on building complex deep learning structures to learn the spatiotemporal feature of the urban traffic flow, ignoring improving the efficiency of the traffic state estimation. Considering the benefit of the tensor decomposition, we present a novel urban traffic state estimation based on dynamic tensor and Bayesian probabilistic decomposition. Firstly, the real-time traffic speed data are organized in the form of a dynamic tensor which contains the spatiotemporal characteristics of the traffic state. Then, a dynamic tensor Bayesian probabilistic decomposition (DTBPD) approach is built by decomposing the dynamic tensor into the outer product of several vectors. Afterward, the Gibbs sampling method is introduced to calibrate the parameters of the DTBPD models. Finally, the real-world traffic speeds data extracted from online car-hailing trajectories are employed to validate the model performance. Experimental results indicate that the proposed model can greatly reduce computational time while maintaining relatively high accuracy. Meanwhile, the DTBPD model outperforms the state-of-the-art models in terms of both single-step-ahead and multistep-ahead traffic state estimation.