用于ADAS实时目标车辆检测的改进SSD算法  

Improved SSD algorithm for real time target vehicle detection in ADAS

作　　者：焦鑫 杨伟东[1] 刘全周[2] 李占旗[2] 贾鹏飞 JIAO Xin;YANG Weidong;LIU Quanzhou;LI Zhanqi;JIA Pengfei(School of Mechanical Engineering,Hebei University of Technology,Tianjin 300130,China;China Automotive Technology and Research Center Co.,Ltd,Tianjin 300300,China)

机构地区：[1]河北工业大学机械学院,天津300130 [2]中国汽车技术研究中心有限公司,天津300300

出　　处：《汽车安全与节能学报》2020年第3期337-344,共8页Journal of Automotive Safety and Energy

基　　金：天津市科技计划项目(18YFCZZC00150);天津市科技计划项目(17YDLJGX00020)。

摘　　要：以实际交通场景中存在重叠小目标车辆为重点,为提升汽车辅助驾驶系统(ADAS)对目标车辆检测的准确性,建立了一种实时目标车辆检测改进算法SSD-P。该算法基于2种方法:1)通过增加小目标特征的提取数量,提出了一种浅层特征图像分辨率重建的方法;2)在非极大抑制中嵌入特征向量进行二次判定方法,以克服单发多盒探测器(SSD)算法对小目标检测精度不高、重叠目标检测能力弱的问题。在PASCAL VOC2012数据集、虚拟交通场景以及实际交通场景中,进行了相关实验验证。结果表明:用该SSD-P算法进行目标车辆检测的平均精度(mAP)为92.4%,比改进前的SSD算法精度提升了4.8%。因此,该改进算法能够改善ADAS的准确性。An improved real-time target-vehicles detection algorithm SSD-P was developed focusing on overlapping small target-vehicles in actual traffic scenes to improve the detection accuracy using an advanced driver assistance system(ADAS).This algorithm was based on two methods:1)a resolution reconstruction method of shallow feature image was proposed by increasing the number of small target feature extraction;2)a quadratic determination method with embedded feature vector in non-maximal suppression to overcome the problems such as low precision and weak ability of overlapping target detection in a single shot multi-box detector(SSD)algorithm.Experiments were carried out in PASCAL VOC2012 data set,virtual traffic scene and real traffic scene.The results show that the mean accuracy precision(mAP)for target vehicle detection is 92.4%by using SSD-P algorithm,this is 4.8%higher than that by using the original SSD algorithm.Therefore,the SSD-P algorithm can improve the accuracy of ADAS.

关 键 词：汽车辅助驾驶系统(ADAS) 实时车辆检测 单发多盒探测器(SSD)算法 小目标 重叠目标 

分 类 号：U461.91[机械工程—车辆工程]