基于YOLO-TridentNet的车辆检测方法  

Vehicle Detection Based on YOLO-TridentNet

作　　者：朱茂桃[1] 邢浩 方瑞华 ZHU Maotao;XING Hao;FANG Ruihua(Jiangsu University,Zhenjiang 212013,China;Shanghai Ganxiang Automobile Mirror Industry,Shanghai 201518,China)

机构地区：[1]江苏大学,江苏镇江212013 [2]上海干巷车镜实业有限公司,上海201518

出　　处：《重庆理工大学学报（自然科学）》2020年第11期1-8,共8页Journal of Chongqing University of Technology：Natural Science

基　　金：上海市科技人才计划项目(16XD1420900)。

摘　　要：针对YOLOv3算法在检测小尺度目标和较远处车辆时存在漏检现象的问题,在分析该算法检测原理的基础上,对其网络结构进行了修改,提出了一种YOLO-TridentNet网络,将三个平行分支网络进行了参数共享;基于KITTI数据集,训练了YOLOv3和YOLO-TridentNet车辆检测模型,比较了两个模型的查准率、查全率、平均精度和每秒检测帧数,验证了模型的有效性。实验结果表明:相比YOLOv3算法,基于YOLO-TridentNet的车辆检测算法的查全率提高了2.2%,达76.5%;查准率下降了0.4%,为98.3%;模型检测速度稳定在15帧/s,下降了2帧;同时,在阈值为0.5的情况下,平均精度为92.35%,提高了1.06%,说明YOLO-TridentNet车辆检测模型能够改善小尺度目标的检测精度。Aiming at the problem of miss detection when detecting small objects and distant traffic by YOLOv3,an improved network named YOLO-TridentNet was proposed on the bias of analysing YOLOv3 algorithm with three parallel branches sharing the same parameters.Based on KITTI dataset,the YOLOv3 and YOLO-TridentNet vehicle detection model were constructed.The precision,recall,average precision and frames per second were respectively compared.The results show that in contrast with YOLOv3,the recall ratio of YOLO-TridentNet increases by 2.2% to 76.5% while the precision decreases only by 0.4% to 98.3%.The detection speed is stable at 15 frames per second reducing by merely 2 frames.Meanwhile,the average precision increases by 1.06% to 92.35%.The results above indicate that YOLO-TridentNe vehicle detection model can improve detection accuracy for objects of small scales.

关 键 词：车辆检测 YOLOv3算法 YOLO-TridentNet模型 小尺度目标 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构] U471.15[自动化与计算机技术—计算机科学与技术]