Sparse Bayesian Learning Assisted Approaches for Road Network Traffic State Estimation

Abstract

Real-time (online) road network traffic state estimation plays a vital role in enhancing the services offered by Intelligent Transportation Systems (ITS). Spatio-temporal data vacancies of contemporary acquisition systems considerably limit the reliability of online traffic state estimation. Online estimation insists smaller temporal frame widths inhibiting the accuracies of low rank matrix reconstruction approaches, while matrix imputation methods do not capture the non-trivial traffic data relationships. Alternatively, this paper investigates traffic state estimation by constructing sparse representations via Sparse Bayesian Learning (SBL) and Block SBL (BSBL) approaches to accommodate under-sampled data, independent of the acquisition type. Appropriate kernel matrices are determined by leveraging historical spatio-temporal correlations among the road network traffic data. Subsequently, unavailable traffic states are estimated from predictive distributions. The estimates are further pruned by kalman filter that corroborates online processing. With the SBL approach, experiments on PeMS traffic data demonstrate less than 6% Normalized Mean Absolute Error (NMAE) for a Signal Integrity (SI) of 0.5. Compared to the state-of-the-art approaches, this value is significantly better. BSBL approach gives similar error performance as SBL despite a SI of 0.26, at the cost of increased computational time. The NMAE from kalman filtered SBL (SBL+K) approach is less than 3.5% in contrast to 4.5% from kalman filtered BSBL (KBSBL) approach, thus demonstrating SBL+K approach as a good compromise between NMAE and computational time, facilitating more accurate online traffic state estimation.