Deep Learning Based Urban Post-Accidental Congestion Prediction

Abstract

Urban roads tend to cause traffic congestion for a long time after the occurrence of traffic accidents, which greatly affects daily transportations. Therefore, the prediction of the duration of traffic jams caused by traffic accidents can allocate traffic resources more reasonably and effectively, release induced traffic information, avoid secondary congestion, and quickly handle traffic accidents. It is of great significance to the rapid rescue of traffic accidents and to eliminate traffic safety hazards. In response to this hot issue, many scholars have done a lot of researches through numerous models, such as probability distribution and time series, and artificial neural networks. However, these models usually only consider temporal features or are based on shallow networks. Therefore, this work adopts a hybrid deep spatial-temporal residual neural network HD-SP-ResNet to predict the traffic volume and velocity, as well as the road congestion duration after traffic accident, so as to monitor and dispatch real-time traffic, response to the postaccidental congestion in time, in order to reduce the various losses incurred by congestion and improve people's satisfaction with traffic on the road. To verify the effectiveness of the proposed model, we conduct extensive experiments based on the taxi trajectory data and road accident data in Shanghai. The experiment results show that the proposed model can achieve a relatively accurate prediction on traffic volume and velocity, as well as the post-accidental congestion duration.