计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (1): 117-125.DOI: 10.3778/j.issn.1002-8331.2206-0316

• YOLO的改进及应用专题 • 上一篇    下一篇

基于YOLOv5的道路目标检测算法研究

王鹏,王玉林,焦博文,王洪昌,于奕轩   

  1. 青岛大学 机电工程学院,山东 青岛 266071
  • 出版日期:2023-01-01 发布日期:2023-01-01

Research on Road Target Detection Algorithm Based on YOLOv5

WANG Peng, WANG Yulin, JIAO Bowen, WANG Hongchang, YU Yixuan   

  1. College of Mechanical and Electrical Engineering, Qingdao University, Qingdao, Shandong 266071, China
  • Online:2023-01-01 Published:2023-01-01

摘要: 为提高道路目标检测精度,基于YOLOv5网络模型,引入自底向上的PANet网络结构,以增强特征融合;采用具有方向感知与位置信息的目标注意力机制,以增强对目标位置的感知能力;增加了一个YOLO检测头,以增强对小目标的学习能力。采用改进的CIOU(ICIOU)目标回归损失函数,使得整个模型对图像特征的学习能力和目标检测精度显著提升。实验结果表明,该模型在华为SODA10M数据集下的mAP达到了68.2%,相比原YOLOv5网络mAP提升了15.4个百分点,检测精度得到了明显提升。在此基础上,对图像尺寸对检测时间和精度的影响进行探索,结果表明适当增大图像输入尺寸,可以在检测速度下降不大(23.3个百分点)的前提下,使得mAP明显提升(3.8个百分点)。

关键词: 深度学习, 目标检测, 注意力机制, 交并比, PANet网络结构

Abstract: In order to improve the accuracy of road target detection, based on the YOLOv5 network model, this paper introduces a bottom-up PANet network structure to enhance feature fusion, adopts a target attention mechanism with direction awareness and location information to enhance the perception of the target position, and a YOLO detection head is added to enhance the learning ability of small targets. The improved CIOU(ICIOU) target regression loss function is adopted, the learning ability of the entire model for image features and the target detection accuracy are significantly improved. Experimental results show that the mAP of this model under the Huawei SODA10M dataset has reached 68.2%, which is 15.4 percentage points higher than the original YOLOv5 network mAP, and the detection accuracy has been significantly improved. On this basis, the paper explores the influence of image size on detection time and accuracy. The results show that appropriately increasing the image input size can significantly improve mAP(3.8?percentage points) on the premise that the detection speed is not significantly reduced(23.3 percentage points).

Key words: deep learning, object detection, attention mechanism, intersection over union, PANet network structure