ST-IFC: efficient spatial-temporal inception fully connected network for citywide crowd flow prediction

Traffic flow prediction is important to urban management for the development of smart cities as well as further contribution to public safety. This paper analysed the spatial and temporal characteristics of large amounts of traffic data in depth and proposed an efficient spatial-temporal inception fully connected (ST-IFC) network for citywide traffic prediction. An inception fully connected (IFC) unit was proposed to capture the spatial dependence and multi-scale characteristics of the traffic dataset. In addition, a multi-level feature fusion strategy is proposed to effectively combine the flow features of low-level surface and high-level abstract to avoid feature loss. The strategy greatly enhances the utilisation of computing resources while improving the prediction results significantly. The simulations were carried out using the trajectory data of Beijing taxis and New York City bicycles. The experimental results show the state-of-the-art performance of our model in addition to high computational efficiency.