ST-DCN: A Spatial-Temporal Densely Connected Networks for Crowd Flow Prediction

Abstract

The accurate prediction of crowd flow is of great significance for urban traffic management and public safety. Its key challenge lies in how to model the complex non-linear spatial-temporal dependencies and other external factors such as holidays and weather conditions. In this paper, we propose a novel deep-learning-based approach to address this problem, called Spatial-Temporal Densely Connected Networks (ST-DCN), which is able to predict both inflow and outflow of crowds in every region of a city. Specifically, ST-DCN consists of three parts: spatial module, temporal module and external module. The spatial module is designed with a densely connected convolutional structure to capture the spatial dependencies at a citywide level. The temporal module is composed of ConvLSTM units to learn long-term temporal dependencies. We propose an external module consisting of fully connected layers for modeling the external factors. Then the outputs of these three modules are merged to predict the final crowd flow in each region. ST-DCN can alleviate the vanishing-gradient problem and strengthen the propagation of spatial features in very deep network. In addition, the spatial features structure can be maintained throughout the network to avoid losing implied spatial information of crowd flow. Experimental results on two real-world datasets demonstrate that ST-DCN achieves significant improvements over the state-of-the-art methods.