Research On Crash Risk Of Urban Road Network Based On Multi-source Data

With the rapid development of urbanization and motorization in modern economic society,the urban city have suffered from severe traffic safety problems,resulting in a new source of growth in road traffic crashes.Urban traffic crashes not only seriously threaten individual lives and property,but also have become the main cause of traffic congestion and road network disorder.Recently,many developed countries in Europe and America have paid more and more attention to the safety studies of urban road network.However,the current urban road network safety studies have suffered from several limitations.First,the traffic exposure are mainly collected from sparse sensors in urban areas and accordingly cannot reflect the actual traffic condition,leading to the low prediction accuracy of crash models.Second,the impact of different types of trips on road network safety have been largely neglected,which greatly restrict the accuracy of regional traffic control.Third,the influence variables usually have significant spatial autocorrelation.Many previous studies usually neglect this point,which greatly restricts the explanatory ability and predictive performance of the crash model.Fourth,existing zonal safety studies usually evaluate the safety level in a long-term period from the average perspective.This static evaluating method cannot dynamically reflect the real-time crash risks.The advancement of information technology results in massive and heterogeneous multi-sourced transportation data every day.For example,data from mobile phone,floating vehicle and smart card have provided us the locations and activity information of travelers,which could be potentially used to depict the traffic condition in urban city.This study was sponsored by the National Natural Science Foundation of China(No.51322810).This study extracted the human activity and trip pattern information from multi-sourced big data.Then,we built the road network crash model to establish the relationship among travel pattern,road network,socio-economic and demographic characteristics and traffic crash risk with the consideration of spatial autocorrelation of explanatory variables.Moreover,this study developed short-term crash risk prediction model,which considers both the spatial and temporal correlation of crash risk in neighboring regions.The developed model could reveal the evolution mechanism of crash risk dynamically,resulting in safety-oriented traffic planning and control in urban city.More specifically,this study includes the following contents:First,this study illustrated the contributing factors to crash occurrence of road network in urban city,including traffic exposure,road network,socio-economic and demographic,and environment.We discussed the traffic exposure variables which are commonly used in previous studies and their limitations.Moreover,we also discussed the feasibility of extracting travel information from multi-sourced big data.Second,this study explored how to extract travel related information from emerging multi-sourced big data.More specifically,this study discussed the procedure of identifying human activities from social media check-in data.Moreover,this study also proposed a novel algorithm for twitter user's potential home location inferring and the calculation method of radius of gyration for each twitter user.Then,this study collected taxi trip data and land use data from New York City and applied Latent Dirichlet Allocation method to discover hidden trip patterns.In addition,this study used random forest method to rank the importance of each trip pattern and select the critical trip patterns to road network level traffic safety.Finally,this study also collected the subway station turnstile sensor data and estimated the distributed subway trips in each region by Kernel Density Evaluation method.Third,this study discussed the spatial autocorrelation among explanatory variables in road network level crash modeling.Then,this study applied the geographically weighted Poisson regression method to develop the crash model.We compared the spatial and temporal patterns between the extracted travel information and traffic crashes,and discussed the relationship between different human activity or trip patterns and crash occurrences.Moreover,this study compared the crash models with various fusion schemes of transportation big data sources.The results suggested that the combined multi-source datasets could complement with each other in the spatial and users coverage when modeling spatially aggregated crash data and accordingly have great potential to increase the confidence and robustness of crash models.Finally,this study also discussed the impact of the sample bias in single big data source on crash models.Fourth,this study developed road network level short-term crash risk prediction model based on deep learning technology.This study proposed a novel spatio-temporal deep learning neural network,which integrates CNN,LSTM and ConvLSTM into one end-to-end architecture.This study compared the hourly crash risk model,daily crash risk model and weekly crash risk model under three kinds of spatial grids,such as 8×3,15×5,and 30×10 grids.The results indicated that the proposed spatiotemporal deep learning approach performs better at capturing the spatiotemporal characteristics for the citywide short-term crash risk prediction in terms of highest prediction accuracy and lowest false alarm rate.Finally,this study collected domestic data to apply the proposed road network level crash model.We used Beijing city as a case study.More specifically,we used Weibo check-in data to extract human activity information.Then,we collected a weekly taxi trip data among the central area of Beijing city.We also collected the land use characteristics to extract the critical trip patterns in China.Finally,we used the spatiotemporal deep learning neural network which has been trained before,to predict the crash risk in the central area of Beijing city.In addition,we also discussed the evolutional pattern and propagation mechanism of crash risk in Beijing.