摘要
针对目前自动驾驶场景下交通目标检测算法抗复杂背景干扰能力弱,导致检测性能不足的问题,提出了一种改进YOLOv5s的复杂道路交通目标检测算法。首先,在特征提取区域,采用多头自注意残差模块(MHSARM)来强化待检目标特征信息,弱化复杂背景干扰;其次,在特征融合区域,采用CoordConv代替传统Conv,使网络具备空间信息感知能力,提升网络检测精度。在开源数据集Kitti及BDD100K上的实验结果表明:改进YOLOv5s算法在复杂道路中具备更强的特征提取能力及良好的泛化能力,mAP_0.5分别达到93.3%和47.4%,与YOLOv5s相比,分别提升了0.9%和1.4%。另外,改进YOLOv5s相较于目前最新的目标检测算法YOLOv7、YOLOv8,mAP_0.5分别提高了1.3%和2.2%,与在Kitti数据集上最新的研究成果Sim-YOLOv4算法相比,mAP_0.5提高了2.2%。
A complex road traffic object detection algorithm was proposed to address the issue of traffic target detection algorithms′inability to resist complex background interference and insufficient detection performance in the current autonomous driving scenario.At first,the multi-head self-attention residual module(MHSARM)was used to improve the feature information of the target to be inspected while decreasing the complex background interference.Secondly,in the feature fusion area,CoordConv was used instead of traditional Conv,so that the network could perceive spatial information and improve network detection accuracy.The improved YOLOv5s algorithm had stronger feature extraction ability and good generalisation ability in complex roads,and mAP_0.5 reached 93.3%and 47.4%,respectively,which was higher than that of YOLOv5s 0.9%and 1.4%.In addition,compared with the latest target detection algorithms YOLOv7 and YOLOv8,the mAP_0.5 of improved YOLOv5s improved by 1.3%and 2.2%,respectively.Compared with the latest research results of Sim-YOLOv4 algorithm on Kitti dataset,mAP_0.5 improved 2.2%.
作者
汤林东
云利军
罗瑞林
卢琳
TANG Lindong;YUN Lijun;LUO Ruilin;LU Lin(College of Information,Yunnan Normal University,Kunming 650500,China;Yunnan Provincial Department of Education Computer Vision and Intelligent Control Technology Engineering Research Center,Yunnan Normal University,Kunming 650500,China;Yunnan Tobacco Leaf Company,Kunming 650500,China)
出处
《郑州大学学报（工学版）》
CAS
北大核心
2024年第3期64-71,共8页
Journal of Zhengzhou University（Engineering Science）
基金
国家自然科学基金资助项目(62265017)。
作者简介
通信作者:云利军(1973-),男,内蒙古呼和浩特人,云南师范大学教授,博士,博士生导师,主要从事视频图像处理、计算机视觉、深度学习算法研究,E-mail:yunlijun@ynnu.edu.cn。
