Spatial-temporal Graph Convolutional Neural Networks For Traffic Accident Forecasting

With the national infrastructure is gradually complete,and the per capita vehicle ownership and the number of traffic participants are also increasing.At the same time,the problem of traffic safety has become increasingly prominent.The occurrence of accidents not only threatens people’s life safety,but also causes the loss of social property.Traffic accident risk prediction can help urban managers effectively and pertinently allocate human,material and other resources to alleviate traffic pressure and prevent traffic accidents.Therefore,the research on traffic accident risk prediction has practical and social significance.At present,with the wide popularization of intelligent sensors and the advent of the digital era,people build intelligent transportation systems through many technologies such as the Internet of things and big data analysis.After the systematic and comprehensive collection of road traffic information indicators,we can dynamically monitor the urban traffic flow,implement intelligent dispatching,improve the efficiency of urban road traffic and alleviate the traffic pressure.In addition,it can also mine the temporal and Spatial Laws hidden behind the traffic accident data to predict the risk probability of accidents.Therefore,this paper takes the traffic related data of New York City as the research object,and forecasts the risk of urban road traffic accidents with the help of classical machine learning and spatio-temporal map convolution depth network model.The main practical work of this paper is as follows:(1)Collection,analysis and preprocessing of multi-source heterogeneous data sets.Firstly,collect the data sets used in this paper,including traffic accident data,traffic flow speed data,weather data,road design data and point of interest data,and briefly introduce and explain the data sets.Secondly,it preliminarily explores the impact of various variables on the risk level of traffic accidents.Then,the heterogeneous data sets are preprocessed,including the unification and transformation of variable units,the correspondence of data in time and space dimensions,the processing of default and abnormal values,feature extraction and so on.Finally,the data processing methods and steps are summarized to form a complete flow scheme of traffic accident data preprocessing.(2)A prediction model of spatiotemporal map convolution neural network is proposed.The model is proposed to solve the problem of traffic accident risk prediction of urban roads.Explain the structural characteristics and construction of the model.The structure of the model adds two convolution layers: space layer and space-time layer,which mainly includes three parts: the first part,spatial map convolution layer.The graph structure is used to simulate the spatial relationship between road sections,and the graph volume product is used to capture the correlation of road space,so as to aggregate the spatial structure feature information of road sections and their neighbors;The second part is time convolution.In the time dimension,the standard convolution layer is further used to update the signal of nodes by merging the information of continuous time nodes.Fitting the relationship between the risk level of the road at the future time node and the historical timestamp;The third part is the full connection layer.A feature fusion module is designed.After processing the external features such as meteorological features and date features,it is connected with other dimensional feature variables as the input of the final prediction.(3)Empirical analysis.In order to verify the effectiveness of the proposed model,this paper takes the data related to traffic accidents in New York as an example,and compares the performance of classical machine learning algorithm and spatio-temporal graph convolution model.The empirical conclusion shows that the performance of spatio-temporal graph convolution neural model is better than the traditional machine learning model in various evaluation indexes.In addition,after removing the three feature modules of space,time and external features,the influence of different dimension features on the performance of the model is compared longitudinally.The results show that the prediction performance of the model without any feature dimension has a certain loss,which is mainly manifested in that the spatial and external features have the greatest impact on the accuracy of the model.The innovations of this paper are summarized as follows:(1)establish the graph structure based on the road network structure data.Convolution is more effective to obtain the spatial feature correlation of roads.(2)Considering multi-dimensional factors,such as POI interest point factor describing the characteristics of road spatial environment,road traffic flow velocity,meteorological characteristics,calendar characteristics and other external factors in time dimension.(3)Construct spatio-temporal map convolution neural network.The main three parts of the model are used to capture the spatial correlation,spatio-temporal dependence and the influence of external factors of roads respectively,Make the accuracy of road traffic prediction higher.