Deep Architecture for Citywide Travel Time Estimation Incorporating Contextual Information

Abstract

To meet the growing demand of accurate and reliable travel time information in intelligent transportation systems, this a develops a deep architecture incorporating contextual information to estimate travel time in urban road network from a citywide perspective. First, several categories of features that affect travel time significantly are analyzed and extracted. On this basis, a deep architecture, which utilizes sparse denoising auto-encoders as building blocks, is proposed to learn the feature representations for travel time estimation. To train the deep architecture successfully, a greedy layer-wise semi-unsupervised learning algorithm is devised. The proposed approach inherently incorporates both the geographical features and contextual features, and accounts for the spatial correlation of adjacent road segments. It is a deep architecture with powerful modeling capabilities for the complex nonlinear phenomena in transportation. The information contained in the huge amount of unlabeled data are fully extracted and utilized. The feature representations for estimation are adaptively learned layer by layer from the input with an unsupervised fashion. The proposed model is applied to the real case study of the road network in Beijing, China, based on the large-scale GPS trajectories collected from a sample of taxicabs. Empirical results on extensive experiments demonstrate that the novel deep architecture provides a promising and robust approach for citywide travel time estimation, and outperforms the competing methods.