Spatio‐temporal expand‐and‐squeeze networks for crowd flow prediction in metropolis

Abstract

The use of deep learning methods to predict traffic flow in transportation systems has become a hot research project. The existing predictive model method faces problems such as long calculation time and difficult data pre-processing, especially for the prediction effect of high traffic area. In this study, the authors propose a novel framework ST-ESNet, spatio-temporal expand-and-squeeze networks, that designs several effective strategies for considering the complexity, non-linearity and uncertainty of traffic flow, and better captures traffic flow characteristics to adapt to the dynamic characteristics of traffic trajectory, traffic duration and traffic flow. Specially, we use extend-and-squeeze process rather than squeeze-and-extend process during the normal residual unit to capture farther spatial dependence among regions. Specifically, inverted residual and deformed convolution structures are utilised in the expanding process, and the convolution with stride 2 is utilised in the squeeze process. Furthermore, image feature scaling is used in each residual unit to obtain more fine-grained surface information, which improves the ability of the model to capture dynamic spatial dependence features. Finally, they use stochastic weight averaging to obtain an integration model. In summary, they propose a new predictive model ST-ESNet. The experimental results show that the authors' proposed network model has better prediction performance compared with the state-of-the-art model.