Leveraging 3D-Raster-Images and DeepCNN with Multi-source Urban Sensing Data for Traffic Congestion Prediction

Abstract

Nowadays, heavy traffic congestion has become an emerging challenge in major cities, which should be tackled urgently. Building an effective traffic congestion predictive system would alleviate its impacts. Since the transit of vehicles heavily depends on its spatial-temporal correlations and effects of exogenous factors such as rain and accidents, they should be simultaneously considered. This study proposes a deep learning approach based on 3D-CNN to utilize many urban sensing data sources wrapped into 3D-Raster-Images. Armed with this, the spatial and temporal dependencies of the data can be entirely preserved. Furthermore, traffic congestion status of different geographical scales at various time horizons can be fully explored and analyzed. We also propose data fusion techniques to (1) fuse many environmental factors that affect vehicles’ movements, and (2) incorporate social networking data to improve predictive performance further. The experiments are performed using a dataset containing four sources of urban sensing data collected in Kobe City, Japan, from 2014–2015. The results show that the predictive accuracy of our models improves significantly when using multiple urban sensing data sources. Finally, to encourage further research, we publish the source code of this study at https://github.com/thanhnn-uit-13/Fusion-3DCNN-Traffic-congestion.