SDCAE: Stack Denoising Convolutional Autoencoder Model for Accident Risk Prediction Via Traffic Big Data

Traffic accident is considered as one of main causes for traffic congestion in cities. There are many causal factors that may give rise to traffic accidents, e.g. driver characteristics, road conditions, traffic flows and weather conditions, etc. Due to uncertain factors as well as the contingency of accident occurrences, it is very difficult to predict traffic accidents. Many existing works have utilized classical prediction models to predict the risk of accidents on highways or road segments. However, predicting the risk of citywide accidents remains an open issue. To address this problem, we propose SDCAE, a novel Stack Denoise Convolutional Auto-Encoder algorithm to predict the risk of traffic accident in the city-level. First, we divided the city into regions by counting the number of accidents and traffic flows in each region. Second, we employed a deep model of stack denoise convolutional autoencoder which considers spatial dependencies to learn the hidden factors in accidents. Third, we conducted extensive experiments on two real-world cross-domain traffic big datasets from a major city of China for accident risk prediction. Experimental results demonstrate that SDCAE could outperforms five baseline methods.