TrajNet: A Trajectory-Based Deep Learning Model for Traffic Prediction

Abstract

Ridesharing companies such as Ube and DiDi provide ride-hailing services where passengers and drivers are matched via mobile apps. As a result, large amounts of vehicle trajectories and vehicle speed data are collected that can be used for traffic prediction. The recent popularity of graph convolutional networks (GCNs) has opened up new possibilities for real-time traffic prediction and many GCN-based models have been proposed to capture the spatial correlation on the urban road network. However, the graph-based approaches fail to capture the intricate dependencies of consecutive road segments that are well captured by trajectories. Instead of proposing yet another GCN-based model for traffic prediction, we propose a novel deep learning model that treats vehicle trajectories as first-class citizens. Our model, called TrajNet, captures the spatial dependency of traffic flow by propagating information along real trajectories. To improve training efficiency, we organize the multiple trajectories in a batch used for training with a trie structure, to reuse shared computation. TrajNet uses a spatial attention mechanism to adaptively capture the dynamic correlations between different road segments, and dilated causal convolution to capture long-range temporal dependency. We also resolve the inconsistency between the fine-grained road segment coverage by trajectories, and the ground-truth traffic data that are coarse-grained, following a trajectory-based refinement framework. Extensive experiments on real traffic datasets validate the performance superiority of TrajNet over the state-of-the-art GCN-based models.