An Intelligent Adaptive Spatiotemporal Graph Approach for GPS-Data-Based Travel-Time Estimation

Abstract

Existing real-world travel-time estimation applications face the crucial challenge of inferencing spatiotemporal traffic status propagation over complex and irregular networks. To meet that challenge, this article presents a novel GPS-data-based travel-time estimation approach utilizing an adaptive spatiotemporal graph (ASTG). The proposed ASTG approach improves the original attention mechanism (with an enhanced self-attention mechanism) while adaptively analyzing the dynamic relevancy between segments over the vast spatial and temporal dimensions. Moreover, various traffic metadata, such as additionally available traffic state variables and network/road characteristic information, were better utilized. Leveraging a gate fusion function, the spatial and temporal dependencies extracted from traffic metadata were fused for inferencing more precise traffic states. A field implementation of the proposed approach was conducted in Zhangzhou, China, with sparse GPS probe data, and evaluated against the automatic vehicle identification reported segment travel time. Compared to other high-performance baseline algorithms, the proposed ASTG model demonstrated state-of-the-art accuracy while intuitively capturing the sophisticated dynamic spatiotemporal relevancy with the proposed enhanced attention mechanism. Implementing the proposed system would provide valuable travel information for road users and, meanwhile, assist traffic management agencies in congestion alleviation.