自动驾驶系统并行加速测试方法研究

Study on Parallel Accelerated Testing Method for Automated Driving System

为提高作为自动驾驶系统安全性验证手段的场景测试的效率,本文中提出了一种自动驾驶系统并行加速测试方法。首先根据自动驾驶系统测试需求,构建了由顶层管理层、中层协调层和底层执行层构成的3层加速测试架构。然后确定不同层级的具体任务:顶层管理单元流动,中层协调参数计算,底层执行试验运行。最后以前车切入场景为例,对某黑盒自动驾驶算法分别使用并行加速、并行遍历、单线程加速和单线程遍历4种方式进行对比测试。结果表明,在测试具体场景总数为4590的情况下,4种方式均发现所有危险场景,但其耗时差别悬殊,分别为1.3、5.7、22.4和96.2 h。说明本文提出的并行加速测试方法可有效、快速地发现参数空间中的危险场景,提高自动驾驶系统测试效率。

To enhance the efficiency of scene-based test as a safety verification means of automated driving system,a parallel-computing-based accelerated test method for automated driving system is proposed in this paper.Firstly,a parallel accelerated test framework is constructed consisting of top management layer,middle coordination layer and bottom execution layer.Then the specific tasks of different layers are determined:including the middle-tier coordination coverage and the specific test parameters of bottom-tier execution;top layer for the management of unit flow;middle layer for the coordination of parameter calculation;while the bottom layer for the execution of testing.Finally,with a scene of front vehicle cutting-in as an example,a comparative testing with four modes,i.e.parallel accelerated,parallel traversal,single line accelerated and single line traversal is conducted on a black-box automated driving algorithm.The results show that in a condition of totally 4590 testing scenes,all dangerous scenes are found by four modes,but with significant difference in consumed time,being 1.3,5.7,22.4 and 96.2 h respectively,demonstrating that the parallel accelerated testing method proposed can effectively and rapidly discover hazardous scenes in parameter space,enhancing the testing efficiency of automated driving system.