Stack ResNet For Short-term Accident Risk Prediction Leveraging Cross-domain Data

With an increasing number of private cars, traffic accident prediction plays an important role in urban management. Existing methods tend to utilize classic learning techniques with historical observations on accident records. Therefore, they fail to consider cross-domain data into consideration and ignore the spatiotemporal dependencies. Recently, as more heterogeneous urban data are available, it is promising to predict citywide accidents in a more fine-grained way. However, it is challenging due to the inherent sparsity of traffic accident records and the complexity of multiple influenced factors. Besides, real-time data cannot be collected completely for short-term forecasting. In our paper, we propose a multi-view spatiotemporal deep learning framework to fuse cross-domain urban data. Specifically, we first assign heterogeneous data into the corresponding grid, then extract the features related to accidents into tensors or vectors. Moreover, we propose a ResNet based multi-task learning framework with speed inference model to realize the prediction of near future's accident risk distribution.