智能汽车环境感知算法测试评价系统开发 ……………………………
隗寒冰 曹 旭 赖 锋
1.重庆交通大学机电与车辆工程学院，重庆， 400074
2.东风汽车集团有限公司技术中心，武汉， 430020
摘要：基于硬件在环仿真平台和六自由度驾驶模拟器，开发了智能汽车环境感知算法测试评价系统。为兼顾仿真测试的可重复性、可拓展性及道路测试的真实性，采用真实交通场景作为输入场景信息；针对现有检测算法评价依赖人工识取的不足，提出了基于机器学习的标准检测算法和基于数据关联的检测评价算法，并将系统测评结果与人工识取结果进行对比。实验结果表明，所开发的测试评价系统方案可行，测试精度较高，实时性较好，在模糊、遮挡、光照变化等复杂环境下均可实现对各种不同特性的车道线和车辆目标的准确测评，能较好地满足智能汽车环境感知算法测试评价的要求。
关键词：硬件在环；六自由度模拟器；环境感知；测试评价；机器学习；数据关联
中图分类号： U461.91
DOI：10.3969/j.issn.1004⁃132X.2018.19.005 开放科学(资源服务)标识码(OSID)： Development of Environmental Perception Algorithm Test Evaluation Systems for Intelligent Vehicles
WEI Hanbing1 CAO Xu1 LAI Feng2
1.School of Mechatronics and Vehicle Engineering，Chongqing Jiaotong University，Chongqing，400074
2.Dongfeng Motor Corporation Technical Center，Wuhan，430030
Abstract： A detection and evaluation system of environment perception algorithm for intelligent ve⁃ hicles was proposed on the basis of hardware ⁃ in ⁃ loop simulation platform and 6 ⁃ DOF driving simulator. The real traffic scenario was implemented as the system inputs in order to take into account both the re⁃ peatability，scalability of virtual tests and the authenticity of road testing together. Aiming at the short⁃ comings of the existing evaluation methods which relied on manual recognitions，a standard object detec⁃ tion algorithm based on machine learning and an evaluation algorithm based on data association were pro⁃ posed. Consequently，the results of the system evaluation were compared with manual recognition ones. Experimental results show that the proposed evaluation system has strong reliability，high precision，and outstanding real ⁃ time performance. The accurate evaluation of lane and vehicle detection algorithm with different characteristics may be realized effectively under complicated road environments such as blur ，oc⁃ clusion and changing illumination. The test evaluation system may meet the requirements of intelligent vehicle environmental perception algorithm.
Key words： hardware ⁃ in ⁃ loop；6 ⁃ DOF simulator；environmental perception；test evaluation；ma⁃ chine learning；data association 0 引言基于机器视觉的环境感知算法是智能汽车的关键和共性技术之一，也是智能汽车领域研究范围最广、水平最深的方向之一。早期车道线和车辆目标检测算法均建立在目标轮廓特征基础上，对检测算法性能测试评价还缺乏足够研究。文献
收稿日期： 2017－04－20
基金项目：国家自然科学基金资助项目（ 51305472）；重庆市自然科学基金资助项目（ cstc2014jcyjA6000） ［ 1 ］用 Sobel算子检测目标纵横向边缘后，再根据尺寸等先验知识进行过滤得到候选边缘，然后经分级过滤、直线型过滤得到最终边缘；文献［ 2 ］利用颜色、形状、直方图特征对候选区域标志牌进行检测。此类方法虽检测效率高，但容易受到光照、阴影、遮挡等因素影响。文献［ 3 ］采用基于特征模板配准方法对目标进行检测，该方法同样受遮挡限制。文献［ 4 ］综述了采用基于目标特征和分类器的目标检测算法以解决光照等环境因素影响的