A Traffic Accident Risk Prediction Model Based on Spatial Gated Memory Network

The prediction of traffic accidents is becoming more and more important in the field of public safety, especially in travel route planning and road safety deployment. Due to the development of deep learning, traffic accident prediction can evolve from traditional machine learning algorithms to neural networks to predict the risk of accidents in a short period of time (hour level) in the future. However, the large urban area and sparse traffic accident samples will bring difficulties to deep learning prediction. If it is not solved, it will easily lead to zero-expansion problems, leading to inaccurate prediction results. In response to the above problems, this paper proposes a space-gated memory network (SGMN) based on deep learning. The model combines real-time accident risk, real-time traffic flow, and real-time weather with multiple inputs to predict sub-regions with high accident risk. In order to verify the performance of the model, two real data sets are used to evaluate the performance of the model. The results show that the performance of SGMN is better than that of commonly used memory neural networks such as RNN, LSTM, GRU, Convulsion and Hereto-ConvLSTM, which validates the model.