Research On Accelerated Generation Method Of Safety-Critical Scenarios For Intelligent Vehicles Based On The Naturalistic Driving Database

Intelligent vehicle represents the strategic commanding heights of future automobile technology.Safety test is the basis and premise of accelerating the industrialization of intelligent vehicle and promoting its large-scale application.At present,intelligent vehicle has developed from the traditional human-vehicle binary system to the human-vehicleenvironment-task coupling system,its complex driving environment makes the safety-critical scenario construction become the research focus in the field of intelligent vehicle test.Firstly,any collision accident may cause great loss of life and property when the intelligent vehicle is driving on the road.In order to ensure the driving safety of the intelligent vehicle,it is necessary to construct safety-critical test scenarios with high coverage.Secondly,the realworld scenarios are characterized by high dimension of scenario parameters,large range of scenarios parameters,and massive scenario combinations.The safety-critical scenarios tend to take up only a fraction of real-world scenarios and have “the long tail effect”.It is urgent to explore the accelerated generation method of safety-critical scenarios to improve the efficiency of intelligent vehicle test.Finally,high confidence test scenario construction is the basic requirement for accelerated generation of safety-critical scenarios.On the one hand,the test scenarios should be reasonable scenarios that intelligent vehicles can encounter when driving on real roads.To avoid meaningless testing process,extreme scenarios that do not exist in the real world should not be generated.On the other hand,the intelligent vehicle test technology should have high confidence and efficiency in the simulation of test scenarios.This paper,which relies on the 13 th Five-Year National Key R&D Program Project "Research on Hardware-in-the-Loop Test Platform Construction and Simulation Test Technology for Automated Vehicles(2018YFB0105103)" and the National Natural Science Foundation of China Project "Simulation Test and Evaluation Method of Automated Vehicles in the Snow and Ice Environment(U22A20247)",researches the accelerated generation method of safety-critical scenarios for intelligent vehicles based on the naturalistic driving database.The naturalistic driving database,which records the vehicle movement states in the traffic environments,is used as the scenario source.Through the abstraction,quantification,segmentation,clustering and derivation of the natural driving database,a high coverage test scenario primitive database is constructed.Combining the scenario primitive reconstruction method and safety-critical scenario optimization search method,the safety-critical scenario set with high coverage,high efficiency and high confidence is construced to improve the efficiency of intelligent vehicle testing and promote the landing of intelligent vehicle industry.This paper revolves around two key scientific problems: construction of high coverage test scenario primitive library based on the naturalistic driving database;generation safetycritical scenarios with high coverage,high efficiency and high confidence.The following main studies have been conducted:Firstly,the construction method of basic test scenario primitive library based on naturalistic driving database is studied.Taking the High D naturalistic driving database as the scenario source,the functional scenarios are defined according to the test requirements,and the natural driving data is abstracted and quantified to generate test scenario sequences.The scenario primitive extraction method based on disentangled sticky hierarchical Dirichlet process hidden Markov model is established to realize non-parametric and interpretable extraction of scenario primitives from scenario sequences.Focusing on the car following scenario primitives and cut-in scenario primitives,an clustering method of scenario primitives based on K-SHAPE is proposed to construct basic test scenario primitive library.Secondly,the scenario primitive derivation method based on generative adversarial network is studied.The conditional generative adversarial network model with gradient penalty based on Wasserstein distance is constructed.In the model,the discriminator network which combines gated recurrent unit and fully connected layer is designed,and the generator network which combines encoder-decoder architecture,attention mechanism and gated recurrent unit is designed.New scenario primitives are derived under the control of the simple scenario constraint,and the test scenario primitives with aggressive driving behavior are generated.Therefore,the diversity and coverage of scenario primitive database are improved.Thirdly,the construction method of the dangerous logic scenario with potential collision risk is studied.The scenario primitive splicing and reconstruction is abstracted as the Markov decision process.The reconstruction method of typical safety-critical scenarios based on twin delayed deep deterministic policy gradient reinforcement learning is proposed.For the car following scenario and cut-in scenario,the corresponding environment state,action space and reward function are designed respectively,and the reinforcement learning models are trained to generate typical safety-critical scenarios.The vehicle motion states in typical safety-critical scenarios are analyzed.The dimension of the logical scenario is reduced,the range of logical scenario parameter is compressed.Finally,the dangerous logical scenario with potential collision risk is constructed to improve the exposure rate of safety-critical scenarios.Fourthly,the safety-critical scenario accelerated generation method in the logic scenario space is studied.The global search method for safety-critical scenarios based on social cognitive optimization algorithm is proposed,and relevant imitation learning rules,observation learning rules,neighborhood search rules and knowledge base updating rules are designed to achieve efficient search for all safety-critical scenario clusters.The local search method for safety-critical scenarios based on the convolution algorithm is proposed.Through scenario mapping and iterative convolution,the remaining safety-critical scenarios in the logic scenario space are searched quickly.The safety-critical scenarios are generated in the car following logical scenario and cut-in logical scenario,and the results show that the safetycritical scenario accelerated generation method can generate safety-critical scenarios with high efficiency and coverage in the logical scenario space.Finally,the hardware-in-the-loop test application of accelerated generation method of safety-critical scenarios is studied.The working principle of the millimeter-wave radar is analyzed,and the millimeter-wave echo simulation method is designed.The millimeter-wave radar in-the-loop test platform is built,and the geometric model and power attenuation model of the radar are designed to realize high confidence scenario-based test for the intelligent vehicle function.The safety-critical scenario accelerated generation method is applied to the millimeter-wave radar in-the-loop test for the tested intelligent vehicle function.The safetycritical scenario set is generated with high confidence,high efficiency and high coverage,and the test efficiency of the intelligent vehicle is improved.