基于反事实推断的自动驾驶路测险态场景推演

Autonomous Driving Road Test Risk Scenario Inference Based on Counterfactual Analysis

摘要:

以自动驾驶接管案例数据为基础，辨析接管干预类型及其与场景要素的关系，构建自动驾驶路测信任接管、避险接管及优化接管3类险态场景；通过搭建CARLA?SUMO联合仿真平台，以反事实推断方法还原险态场景，利用路测实际数据训练自动驾驶车辆模型，并推演险态场景下自动驾驶运行状况，从安全、效率、舒适度等维度提出自动驾驶智能行驶水平评价指标。分析表明，反事实推断方法在实现自动驾驶路测险态场景推演方面具有优异表现；同类险态场景下自动驾驶小客车智能行驶水平优于卡车；不同类险态场景的场景要素差异性显著，且相较于其他类型接管场景，避险接管场景自动驾驶智能行驶水平更为可靠。

Abstract:

Based on the data of takeover cases in autonomous driving road testing， this paper analyzes the types of takeover interventions， and sets up three types of risky scenarios： trust， risk-avoidance， and optimization takeovers. By developing a CARLA-SUMO joint simulation platform under the OpenCDA framework and using the counterfactual inference method， the operating conditions of autonomous driving in risky scenarios are deduced from the road-testing datasets. From the perspectives of safety， efficiency， comfort， etc.， multi-dimensional evaluation indicators for the intelligent driving level of autonomous driving are proposed. The results show that the counterfactual inference method is capable of achieving the deduction of risky scenarios under the takeover conditions. In similar risky scenarios， the intelligent driving of autonomous passenger cars is better than that of trucks. The differences in fundamental elements and operational indicators among risky scenarios are significant， and intelligent driving in risk-avoidance takeover scenarios is more reliable compared to trust takeover scenarios.