Research On Traffic Accident Black-spots Identification Method Based On Mixed-type Data Neural Network

Road traffic accident injuries are already one of the top ten causes of death in the world,and globally,as many as 1.25 million people die and 50 million are injured each year.If the locations of traffic accidents can be identified,this will have a huge beneficial effect on reducing the number of traffic accidents.The purpose of this thesis is to identify the traffic accident black-spots.The traditional identification models of traffic accident black-spots are mostly built on the basis of historical traffic accident data for mathematical statistics,clustering or other machine learning identification.These ways lack of consideration of the spatiotemporal sparsity of accident data,and the cost of obtaining data is higher,and the accuracy of identification is lower.For these issues,this thesis intercepts the area of interest of the City of London and establishes a spatial grid system.Google Maps Static API is used to capture satellite images of these grids and input them into CNN.Extract image features of accident-prone areas,and then combine key features of traffic accidents with image features extracted from CNNs to create mixed-input or mixed-type data neural networks(MDNN).Both types of data are input into separate deep learning models and their output is combined into the final layer to identify whether a traffic accident has occurred in a given area.The details are as follows:First,by analyzing the different characteristics of traffic accident data,the initial screening of the influencing factors that cause traffic accidents,and then using the road component hotspots cause analysis model based on the contribution of the principal components to further find the main causes of traffic accidents,finally getting the key features of identifying traffic black-spots through Granger causality analysis.Second,a spatial grid system is established for the area of interest.Combining with traffic accident data,the Google Maps Static API captures satellite images of each grid to define a label set to generate a sample set.CNN is applied to the classification of traffic accident black-spots,and established a model by learning the differences between different roads,and identified new road areas.This broadens new areas for CNN applications.Last,models built solely on the basis of structured or image data are expensive and have low accuracy.This thesis proposes a traffic accident black-spots identification model based on MDNN,which using CNN branches to process image data,and using MLP branches to process traffic accident data,and then connecting these branches together to form the final neural network model to obtain better classification results at a lower cost,and a wide range of application prospects.