基于改进YOLOv3的车辆识别方法  

Vehicle Recognition Method Based on Improved YOLOv3 Algorithm

作　　者：王永顺[1] 贾文杰 王晨飞 宋慧 Wang Yongshun;Jia Wenjie;Wang Chenfei;Song Hui(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou,Gan8u 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730070

出　　处：《激光与光电子学进展》2021年第16期232-239,共8页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61366006)。

摘　　要：针对车辆目标检测中存在的小目标检测准确率低、系统鲁棒性差的问题,提出一种改进的YOLOv3算法对车辆进行目标检测。首先,该算法将空洞卷积引入到YOLOv3算法的下采样层,提高特征图的分辨率,加强对小目标的检测效果;其次,针对车辆图像中小目标识别的问题,将YOLOv3的3个检测尺度增加至4个并相互融合不同尺度特征层的信息,改进后的空间金字塔结构实现了对小目标检测进行进一步增强的目标;最后,采用Complete IoU(CIoU)作为损失函数,使目标框回归更加稳定,并且训练中不会出现发散现象。在KITTI数据集上的测试结果表明,所提改进的YOLOv3算法能获得较高的检测精度,平均检测精确度提高了4.6%,检测速度约为44.1 frame/s,在提高精度的前提下仍保持良好的检测速率。An improved YOLOv3 algorithm that detects target vehicles is proposed to address the problems of low detection accuracy of small targets and poor robustness of systems in target vehicle detection.First,the proposed algorithm introduces the dilated convolution into the downsampling layer of the YOLOv3 algorithm,improving the resolution of the feature maps and detection effect of small targets.Second,to address the problem of small target recognition in vehicle images,the proposed algorithm increases the three detection scales of YOLOv3 to four in addition to connecting and fusing the information with different scales,and the improved feature pyramid structure further improves small target detection.Finally,using Complete IoU(CIoU)as the loss function makes the target frame regression more stable,and there is no divergence in training.The KITTI dataset test results show that the improved YOLOv3 algorithm can achieve high detection accuracy.The proposed algorithm improves the average detection accuracy by 4.6%,and the detection rate is approximately 44.1 frame/s.On the premise of improving the accuracy,the proposed algorithm maintains a high detection rate.

关 键 词：图像处理 车辆目标检测 YOLOv3 空洞卷积 尺度检测 损失函数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]