自动驾驶算法的异常事件生成系统设计与评估

评价自动驾驶算法对异常交通事件的响应具有重要的应用价值，针对在真实世界中制造异常交通事件存在代价高、风险大等问题，本文提出了一种面向自动驾驶算法评估的异常交通事件生成方法，该方法可以自动生成5类异常交通事件；基于该方法，搭建了一个驾驶策略评估系统，以实现对自动驾驶算法的评估.为验证生成的异常交通事件的有效性，对一种基于深度网络的模仿学习自动驾驶算法进行了评估.实验结果表明，生成的异常交通事件可以更全面地评估自动驾驶算法的性能. 

The Design and Evaluation of Abnormal Event Generation System for Autonomous Driving Algorithms

Evaluating the response of autonomous driving algorithms to abnormal traffic events acts an important application value. However, generating abnormal traffic events in the real world is costly and risky. In this paper, a method was proposed to generate abnormal traffic events for the evaluation of autopilot algorithms. The method was arranged to automatically generate five kinds of abnormal traffic events. And a Driving Strategy Evaluation System was established to realize the evaluation of the autopilot algorithm. An imitation learning autopilot algorithm based on a deep network was evaluated by the proposed system in order to verify the validity of the generated abnormal traffic events. The experimental results show that the generated abnormal traffic events can more comprehensively evaluate the performance of the autonomous driving algorithm.