Research On Derivation Method Of Longitudinal Collision Avoidance Test Scenarios Based On Natural Driving Data

The mileage based test method has been proved unsuitable for autonomous vehicle development due to its low test efficiency.How to ensure the safety of the autonomous driving system in the design stage and conduct reasonable test evaluation in the test and verification stage has become an urgent problem to be solved.Scenario-based test and evaluation method is widely used in the field of intelligent vehicle test and evaluation due to its flexibility,repeatability,high test safety and low test cost,and has gradually become a normative method of autonomous driving test and evaluation.However,the randomness,uncertainty and selfadaptability of the real traffic result in the infinite running scenes of vehicles.Based on the analysis of expert experience,effective test scenes can be obtained,but it is difficult to ensure the diversity and coverage of scenes.Using data mining method to derive scenes from natural driving data is faced with challenges brought by the black box model of machine learning algorithm to generate scenes in rationality and interpretability.In order to fully explore the unknown reasonable operation scenario of the autonomous driving system,this paper integrated the safety analysis method and the natural driving data set mining learning theory,used the deep learning model and expert experience to explore the test scene space,and conducted the following research on the longitudinal collision avoidance test scenario of the autonomous driving system:Firstly,the HWP system is taken as an example to analyze the longitudinal collision avoidance safety of the automatic driving system,and the knowledge-based longitudinal collision avoidance function scenario is obtained.Define related items of the system to be analyzed,that is,define the functions and hardware architecture of the HWP system,design and operating conditions(ODC),and analyze the initial architecture of the system.Then,based on the HAZOP method,the guide words are used to identify the unexpected behaviors of braking functions involved in the longitudinal collision avoidance scene,and the identified unexpected behaviors are combined with the operation scene to form vehicle-level hazard events.According to the severity(S),exposure rate(E)and controllability(C)evaluation indexes and values recommended by ISO26262,the risk assessment of hazardous events was carried out,and the vehicle safety integrity level(ASIL)of the aforesaid hazardous events was obtained.The knowledge based typical collision avoidance scenario of HWP was obtained by concluding unacceptable hazardous events.It provides a basis for adding road constraints to the scenarioderived learning model and supports the subsequent testing and evaluation of the measurement system.The Argoverse natural driving scene database was used for inference and derivation of collision avoidance scenes,and the longitudinal collision avoidance derived function scenes generated based on data were obtained.In the traditional Variational Auto-Encoder(VAE)model,a GRU gate loop control unit is introduced to correlate the state before and after scene frames,and the features of generative network and cyclic neural network are integrated to realize the learning and generation of time sequence trajectory data.In order to enhance the authenticity of the derived scenes,the Ada IN style transfer method is used to apply the road structure pictures corresponding to the natural driving data as constraints to VAE track generation,and a GRU-VAE track generation model is constructed with integrated road conditions.The model uses Argoverse track data for training,and then uses the track data set generated by the model for clustering after dimensionality reduction by encoder,to obtain the derived automatic driving longitudinal collision avoidance test scene equivalence class,and then inductive longitudinal collision avoidance function scene of semantic.A software in-loop test platform is constructed to realize the performance testing and evaluation application of the autonomous driving system under test in the above typical longitudinal collision avoidance scenarios.Firstly,the longitudinal collision avoidance functional scenarios based on knowledge and data are comprehensively analyzed,typical test scenarios are summarized,and the functional scenarios are parameterized according to laws and regulations,and test cases are designed.Prescan was used to build a virtual test scene,and then Prescan and Simulink were used to build an in-ring test system of the autopilot system software to realize the test evaluation of the autopilot system under the typical longitudinal collision avoidance test scene,which verified the test scene obtained by the scenario mining and analysis method proposed in this paper.