Deep spatio-temporal 3D densenet with multiscale ConvLSTM-Resnet network for citywide traffic flow forecasting

Reliable traffic flow forecasting is paramount in Intelligent Transportation Systems (ITS) as it can effectively improve traffic efficiency and social security. Its vital challenge is to effectively integrate various factors (such as multiple temporal correlations, complex spatial correlation, high heterogeneous) to infer the evolution trend of future traffic flow. Inspired by spatio-temporal prediction in computer vision, we regard traffic data slices at each moment as traffic frames. This paper presents an end-to-end architecture named Spatio-Temporal 3D Densenet Multiscale ConvLSTM-Resnet Network (ST-3DDMCRN) to predict future traffic flow accurately. Specifically, a 3D densenet network is applied simultaneously to capture the traffic frame's local regional spatio-temporal information. Traditional Resnet networks cannot capture long-range spatial correlation, a novel multiscale ConvLSTM-Resnet network is developed to overcome this problem, extracting traffic frame's nearby and long-range spatial dependencies. In addition, considering the spatio-temporal heterogeneity of traffic frames, a Region-Squeeze-and-Excitation (RSE) unit is designed to accurately quantify the difference of the contributions of the correlations in space. The experiment result on two datasets in the real world illustrates the ST-3DDMCRN model outperforms the state-of-art baselines for the citywide traffic flow prediction. Furthermore, to validate the model's generality, we utilize the model to predict the passenger pickupidropoff demand task, the prediction results are more accurate than the baseline methods. (C) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes an end-to-end architecture called ST-3DDMCRN for accurate prediction of future traffic flow by effectively integrating various factors. The architecture utilizes a 3D densenet network to capture the spatio-temporal information of traffic frames, and a multiscale ConvLSTM-Resnet network to extract the spatial dependencies of the frames. The Region-Squeeze-and-Excitation (RSE) unit is designed to accurately quantify the contribution differences of spatial correlations. Experimental results demonstrate that the ST-3DDMCRN model outperforms baseline models in citywide traffic flow prediction and shows good generality in predicting passenger pick-up/drop-off demand.