Research On Identification And Risk Assessment Of Potential Dangerous Scenes On Urban Roads

Data driven deep learning in technology has greatly promoted the process of automatic driving vehicle production.However,the "Long Tail Problem" has been in suspense,which largely restricts the commercialization of the automatic driving vehicle.The potential dangerous traffic scene caused by the visual field occlusion of on-board sensors is a typical"corner case" of "Long Tail Problem".There are two main types of problems in the research of potential dangerous traffic scenes at home and abroad.One is that the research scene is single;Second,motion planning only considers static occlusion blind area.However,in reality,there are many kinds of potential dangerous traffic scenes with different characteristics.To solve these problems,this paper takes the city road condition as background,classifies,identifies and estimates the potential dangerous traffic scenes,and carries out the plan of the speed of the autopilot.Firstly,classification of potentially dangerous traffic scenes.The problem of occlusion of human driving vision is extended to the field of automatic driving.Combined with a large number of traffic scene analysis,the commonness and characteristics of scene features are found.The moving state of occlusion is divided into static occlusion and dynamic occlusion;On this basis,combined with road features,scene commonness and relevant theoretical knowledge,it is divided into three potential dangerous traffic scenes:static gradual change blind area,static abrupt change blind area and dynamic gradual change blind area,and the environment models of each kind of scene are established to reduce the difficulty of subsequent identification module.Secondly,the identification of potential dangerous traffic scenes.According to the features of the scene model,different recognition methods are adopted.Considering the static gradual change blind area and dynamic gradual change blind area,the main research object is vehicles,and the deep learning algorithm uNetXST network is used for recognition.Compared with the deep network and its variants,the F1 score of this model is 92%,and the Mlou score is 87.83,which are higher than the deep network and its variants;Considering the static mutation blind area,the research object is the location of the gap of the isolation zone and the road connecting region,and the region growth method is used to identify the blind area;The effectiveness and robustness of the recognition algorithm are verified through the simulation experiments of structured and unstructured roads in different cities,which provides the starting point of potential dangerous areas for subsequent motion planning.Thirdly,risk assessment and speed planning of potential dangerous scenes.In order to increase the robustness of risk assessment model,dynamic Bayesian network based on scene context information is used for risk assessment.Eight safety factors in road environment are taken as assessment network nodes to increase the flexibility of assessment model;In order to drive safely and efficiently,based on human driver’s behavior,a longitudinal motion planning model is established,and the risk points of occlusion area and road environment information are introduced into the model,so that the longitudinal planning speed changes with the change of relative distance and risk value,so as to reduce the impact on road traffic flow,and increase the robustness of the model.Fourth,the experimental verification of risk assessment and motion planning model of potential dangerous scenes.The campus road scene is selected for real vehicle experiment.Aiming at the static gradual change blind area and static sudden change blind area scenarios,considering a variety of complex conditions and test indicators,the risk assessment model and motion planning model of potential dangerous traffic scenes are tested offline,and the simulation experiment and human driver experiment data are compared and analyzed.The experimental results show that this model can not only ensure the driving safety,but also realize the high efficiency of vehicle traffic flow,and improve the active safety of autonomous driving vehicles.