Research For Traffic Accident Risk Prediction Based On Deep Learning

With the rapid development of urbanization and the realization of road motorization,the lives of the people are more convenient.At the same time,the massive use of motor vehicles puts tremendous pressure on the government’s traffic control,causing a series of social problems such as traffic congestion,air pollution and traffic accidents,especially frequent traffic accidents,leading a huge loss to people’s life safety and social property.The results of traffic accident risk prediction can help city managers to rationally deploy police to relieve traffic and avoid traffic crash.Furthermore,it can provide safety guidance for personal travel.Therefore,accurate and effective prediction of future regional traffic accident risk has important research significance and social value.Due to the widespread use of sensors and the wide application of data collection,people can obtain a large number of multi-source heterogeneous data related to traffic accidents.Instead of relying solely on traffic accidents or traffic flow data,people can study traffic accident risks more comprehensively.This thesis preprocesses multi-source heterogeneous data and uses machine learning and deep learning models to achieve accurate and effective prediction of regional traffic accidents in the city.The main contributions of this thesis include the following aspects:1.First,a series of multi-source heterogeneous data sets related to traffic accidents is introduced,including traffic accident data,travel data of different vehicles,weather data,road design data and point of interest data,etc.,and then their impact on traffic accidents is studied.Secondly,according to the spatio-temporal characteristics of the data,the data is preprocessed accordingly.The preprocess includes spatio-temporal correspondence of data,default value filling,feature extraction and data normalization.Finally,a set of preprocessing procedures for solving traffic accident risk problems is summarized,and all data is processed into a type suitable for traffic accident risk prediction.2.A deep learning framework is presented to predict regional Traffic Accident risk that utilizes a Spatial-Temporal Attention Network(named TA-STAN).The model is based on the traditional encoder-decoder structure,adding a spatial attention mechanism and a temporal attention mechanism.This model consists of three main parts:(1)Spatial attention mechanism.In the encoder phase,two spatial attentions,local spatial attention and global spatial attention,are used respectively,to fit the dynamic impact of local historical various traffic indicators on local future traffic accident risks and the dynamic impact of global historical traffic indicators on local future traffic accident risks.(2)Temporal attention mechanism.In the decoder phase,temporal attention is used to fit the dynamic relationship between predicted values of different future timestamps and historical timestamps.(3)External feature fusion.In the decoder stage,the model incorporates external features,e.g.,weather features,time features and road design features as part of the final prediction.A feature fusion model is designed to more accurately fit the traffic accident risk for a period of time in the future.The experimental results show that the TA-STAN model is superior to the traditional machine learning model in all three evaluation functions.More importantly,by visualizing the weight of attention,the actual meaning of attention weights can reasonably be interpreted,which plays a crucial role in our model.