ST-TAP: A Traffic Accident Prediction Framework Based on Spatio-Temporal Transformer

As an important part of urban management, researchers have made many efforts in accident prediction. The current research mainly considers the impact of the temporal features of traffic flow and the fixed topology of the road on the accident. However, these studies do not consider the changes in the relationship between road spatial features and accidents over time. In order to better achieve the accident prediction, we first take the traffic feature parameters of traditional highways traffic fixed station data after pre-processing as input dataset. Then, we use ClusterGAN to combine traffic spatio-temporal features to cluster the dataset and divide it into multiple datasets. After that, we combine the temporal and spatial features of the traffic flow with the temporal and spatial Transformer to capture the static and dynamic temporal and spatial dependence of the traffic flow, and obtain the accident probability through multiple extractions of temporal and spatial dependence of the traffic flow and convolutional neural network to accurately predict traffic accidents in real time. Finally, we compared with a variety of existing methods to verify the advance of this method.