面向自动驾驶测试的交互场景策略建模与仿真

Interactive Scenarios Strategy Modeling and Simulation for Automated Driving Testing

自动驾驶汽车与人类驾驶汽车的交互能力对未来新型混合交通的运行安全和效率至关重要。为测试高等级自动驾驶汽车的交互能力,测试场景中的背景车须具备自然驾驶交互特征并反映人类驾驶员异质性交互策略。本文基于博弈论框架,建立了驾驶交互策略模型(game-theoretical strategic interaction model, GSIM)。GSIM通过在个体效用函数中引入可差异化取值的交互社会性表征参量,实现背景车交互策略的定向调控。十字路口无保护左转场景的测试实验表明,GSIM可保留自然驾驶逐步规划、双向交互的可解释性,保障交互行为的仿真精度;同时,可有效复现高风险场景中人类驾驶的交互策略,有助于提供具有挑战性的高测试价值场景。对比传统智能驾驶人模型,GSIM模型在无保护左转场景中轨迹仿真精度平均提升42.8%,严重冲突事件复现率提升25.8%。

The interaction ability between Highly Automated Vehicles(HAV)with human-driven vehicles is critical to the operational safety and efficiency of hybrid traffic in future.In order to test the interactivity of HAV,the background vehicle in the testing scenario needs to have naturalistic interaction characteristics and reflect the heterogeneous interaction strategy of human drivers.Based on the game theory,the Game-theoretical Strategic Inter-action Model(GSIM)is developed in this paper.In the individual utility function,the interactive social character-ization parameters with distinguishable values are introduced to directionally regulate the interaction strategy of the background vehicle.The test results of unprotected left-turning scenarios at intersections show that GSIM preserves the interpretability of natural driving stepwise planning and mutual interactions to ensure simulation accuracy of in-teractive behaviors.GSIM is also able to effectively reflect the interactive strategy of human driving in high-risk sce-narios,helping to provide challenging and valuable testing scenarios.Compared to traditional Intelligent Driver Models,GSIM improves average simulation accuracy by 42.8%in unprotected left turn scenarios and serious con-flicts recurrence rate by 25.8%.