摘要
针对行人检测中对小尺度目标和遮挡的检测困难问题,提出一种基于改进YOLOv5的行人目标检测方法。结合GhostNet将YOLOv5的CSP模块改进为CSPGhost模块,对于相似的特征,将复杂的卷积运算简化成线性运算;在每个CSPGhost模块后面插入通道注意力模块,保证了模型检测速度的同时具有较高的检测精度;优化空间金字塔池化层,在不改变原有效果的前提下,降低算法的时间成本;将边框回归损失函数GIoU优化为考虑了长度损失和宽度损失的EIoU,其回归速度更快,得到的回归结果更好。实验结果表明:基于CSPGhost改进的YOLOv5的行人目标检测方法在COCO数据集上种类平均精度值为55.8%,检测速度达到374帧·s^(-1),对小目标的检测能力更强,对遮挡条件下的目标漏检率更低,检测速度更快,能够达到行人检测的实际应用要求.
With the aim of detecting small targets and occlusions in pedestrian detection,a pedestrian detection method was proposed based on improved YOLOv5.In combination with GhostNet,CSP module in YOLOv5 was improved to CSPGhost module,so complex folding operation was simplified for similar functions on linear operation.Channel attention module was used behind each CSPGhost module to ensure model recognition speed and high detection accuracy.The pooling level of spatial pyramid was optimized to reduce time cost of the algorithm without changing original effect.The frame regression loss function GIoU was optimized for EIoU,which took into account the length loss and width loss.Their regression rate was faster and the regression results were better.Experimental results showed that the pedestrian target detection method based on the improved YOLOv5 on the basis of CSPGhost had a mAP value of 55.8%in COCO dataset,and detection speed reached 374 FPS.It had stronger detection capability for small targets,lower error detection rate for targets under occlusion,and faster detection speed,which could meet practical application requirements of pedestrian detection.
作者
谢英红
周育竹
韩晓微
高强
贾旭
XIE Yinghong;ZHOU Yuzhu;HAN Xiaowei;GAO Qiang;JIA Xu(School of Information Engineering,Shenyang University,Shenyang 110044,China;Academy of Science and Technology Innovation,Shenyang University,Shenyang 110044,China;School of Electronics and Information Engineering,Liaoning University of Technology,Jinzhou 121001,China)
出处
《沈阳大学学报(自然科学版)》
CAS
2024年第3期205-212,共8页
Journal of Shenyang University:Natural Science
基金
2020辽宁省教育厅科学研究经费项目(ZGXJ2020010)
辽宁省教育厅面上项目(LJKMZ20221827)
辽宁省应用基础研究计划(2022JH2/101300279)。