Research On Traffic Accident Risk Prediction Via Fusing Multi-source Spatial-temporal Data

In recent years,with the improvement of people’s living standards,the number of cars in the country has increased year by year,which has brought great convenience to people’s travel,and also increased the risk of traffic accidents,posing a huge threat to people’s lives.This paper focuses on predicting the risk of urban traffic accidents based on fusing multi-source spatial-temporal data.If the traffic accident risk can be accurately predicted in advance,the government can make better traffic planning and management measures,and individuals can choose safer roads and time to travel.Traffic accident risk prediction is a very challenging task,and it mainly faces the following three difficulties:(1)The impactors of traffic accidents are complex,since the occurrence of traffic accidents is usually affected by multi-source spatial-temporal factors,such as POI,weather,road conditions,etc.(2)Traffic accidents usually exhibit multiscale dependencies in both the spatial and temporal dimensions,including the geographical spatial-temporal correlations and the semantic spatial-temporal correlations.(3)Traffic accident data exhibits zero inflation issue,since traffic accidents are smallprobability events,which cause too excessive number of zero-values existing in traffic accident risk samples.Existing works cannot solve the above three problems at the same time.Therefore,this paper studies the traffic accident risk prediction based on fusing the traffic accident historical data and other spatial-temporal data.Firstly,this paper proposes a traffic accident risk forecasting network,named STRisk Net,by fusing local and global spatial-temporal features.It models traffic accident risk from both local spatial-temporal correlations and global spatial-temporal correlations.The model is composed of three modules,including local region spatial-temporal correlation module,global region spatial-temporal correlation module and fusion prediction module.In the local region spatial-temporal correlation module,we construct spatial-temporal grid data and utilize CNN to capture the spatial correlations among neighboring regions.Then GRU and attention mechanism are used to model the temporal correlations of traffic accidents.In the global region spatial-temporal correlation module,we construct a road similarity graph to represent the similarity of road features between regions.Then graph convolutional neural network is used to simultaneously capture the correlations between nodes in both spatial and temporal dimensions.The fusion prediction module merges the outputs of the above two modules to obtain the final prediction results.In order to solve the zero-inflated issue,this paper proposes a weighted loss function.When calculating the loss,according to the different accident risk levels,different weights are assigned to the samples.Secondly,this paper further explores the spatial-temporal correlations and improves the ST-Risk Net model.Specifically,we consider the local spatial-temporal correlations among regions from the geographical aspect,while the global spatial-temporal correlations from the semantic aspect.Based on this idea,we propose a traffic accident risk forecasting network,named GSNet,by fusing geographic and semantic spatialtemporal features.The model is composed of three modules,including geographic spatialtemporal correlation module,semantic spatial-temporal correlation module and fusion prediction module.The difference from ST-Risk Net is that in the semantic spatialtemporal correlation module,we construct the risk similarity graph and the POI similarity graph to describe the similarity among regions from different semantic aspects.Then GCN is used to model the spatial correlations between nodes,while GRU and attention mechanism are used to model the temporal correlations of traffic accidents,thereby the model further improves the prediction accuracy.Finally,extensive experiments are conducted on two real-world traffic accident datasets.The experimental results show that both ST-Risk Net and GSNet outperform the existing baseline models,demonstrating they can effectively model the multi-scale spatial-temporal correlations of traffic accident data.In addition,GSNet achieves better performance than ST-Risk Net since it sophisticatedly models the spatial-temporal correlations from both geographical and semantic perspectives.