Online Safety Verification Model For Autonomous Planning Algorithms In Urban Scenarios

Safety is the key to the mass production of autonomous vehicles,which is also the most complicated and urgent difficulty in the development of autonomous vehicles.Functional safety for faulty risks can no longer meet the safety requirements of highly complex systems,the risks caused by system functional limitations have attracted more and more attention.A complete autonomous driving system can be divided into three parts: perception,planning and control.Among them,the planning layer selects appropriate driving behaviors based on many factors such as safety and comfort,then generates the driving trajectories.Designing and verifying a safe and efficient planning algorithm based on perceptual information is the key to autonomous system development.At present,the testing and evaluation methods for the safety of autonomous planning algorithms mainly include simulation testing,closed site testing,actual road testing and auditing.But there are differences between simulation testing and reality,and it is difficult to enumerate all dangerous scenarios in closed site testing.The actual road testing takes long time,which is hard to meet the requirements of development cycle.The system engineering methods for auditing also have corresponding limitations.In order to solve the problems of the above-mentioned,the industry tends to adopt the formal verification.The formal verification method is based on driving common sense to establish accurate mathematical formulas,and form a traceable and verifiable model,which is used to guide vehicles to make safer driving behaviors.This paper formalized the legality and rationality rules of driving vehicles,and proposed corresponding constraints.For the planning algorithms of autonomous driving,an online safety verification model was designed to evaluate and ensure the legality and rationality of driving behavior.Legality means that vehicles can abide by the traffic rules strictly,while rationality means that vehicles can avoid obstacles.As a parallel safety layer with the main planning system,the online safety verification model can check whether the autonomous driving behaviors meet the requirements in real time,and provide braking or steering responses in critical situations.Firstly,the urban world model was constructed to extract the perception and map information required in the safety verification stage,which served as the basis for subsequent safety assessment and response.The urban world model included scene model and dynamic parameter optimization model.The scene model contained road model,traffic signal model,self-vehicle model,and traffic participant model.Meanwhile in order to improve the robustness of the online safety verification model,we also used the typical classification algorithms(Decision Trees,SVM,Ensemble)and the neural network conjugate gradient back propagation algorithm to build the dynamic parameter optimization model,which takes the weather,light,speed,etc.as the input parameters and driving conditions as the output parametersSecondly,the principle of online safety verification was defined.The online safety verification model included modules of safety assessment for planning algorithms and safety response to dangerous state.The safety assessment content was divided into two aspects: legality and rationality of driving behaviors.We formalized legal rules such as traffic signal rules,lane driving rules,speed rules,intersection driving rules,lighting rules,as well as rational rules such as active obstacle avoidance rules,right-of-way rules,then evaluated the dangerous state based on the constraints.The safety response aimed at the dangerous state that violates the legality and rationality,proposed corresponding safety response measures,and provided the final control instruction by conducting response arbitration.Finally,Simulink&Pre Scan was used to complete the simulation verification,while the real vehicle verification was completed based on Haval H7 platform.The simulation verification included two parts: the first part selected real accident cases from the National Automobile Accident In-Depth Investigation System,and tested the effectiveness of the online safety verification model through case repetition.The second part built a continuous real scene according to the public map information,and observed the performance of different planning algorithms under the supervision of the safety verification model.In this paper,the vehicle-vehicle test scenarios of autonomous emergency braking system in China-New Car Assessment Program were reproduced,and the effectiveness of the online safety verification model was tested by comparative tests.The test results showed that the online safety verification model performed well in most cases and could improve the legality and rationality of vehicle driving behaviors effectively.