面向道路交通场景的轻量级目标检测方法  

Lightweight object detection method for road traffic scene

作　　者：黄仝宇 胡斌杰[1] 朱婷婷 HUANG Tongyu;HU Binjie;ZHU Tingting(South China University of Technology,Guangzhou 510640,China;Guangdong Baiyun University,Guangzhou 510450,China)

机构地区：[1]华南理工大学,广东广州510640 [2]广东白云学院,广东广州510450

出　　处：《现代电子技术》2022年第3期88-95,共8页Modern Electronics Technique

基　　金：国家自然科学基金项目(61871193);广东省自然科学基金重点项目(2018B030311049);广东省重点科技领域研发计划(2019B090912001)资助。

摘　　要：针对道路交通场景下的目标检测算法模型占用系统资源较多,对小目标、遮挡目标的检测精度较低等问题,提出一种基于改进的YOLOv5s的轻量级目标检测方法。首先,将主干网络中一些运算量较大的模块替换为Ghost模块或者深度可分离卷积模块,可以减小网络规模、提高推理速度;其次,在主干网络添加SE模块,筛选针对通道的特征信息,提升特征表达能力;再次,使用排斥力损失函数RepulsionLoss作为bbox损失函数,使目标的预测框与匹配的目标框的距离缩小,与周围非匹配目标框的距离加大;然后,采用DIoU⁃NMS作为后处理方法,在抑制准则中不仅分析重叠区域,而且还计算两个框之间的中心点距离,可以提升遮挡情况下目标检测的精度;最后,构建道路交通场景下交通参与者的数据集,共计61200张,其中48960张作为训练集,12240张作为测试集,并在主流的GTX1080 GPU硬件平台进行验证。文中方法的mAP为85.83%,FPS为76.9 f/s,模型大小为25.6 MB,其mAP比YOLOv5s高出0.86%,FPS和模型大小均优于YOLOv4和YOLOv5系列算法。实验结果表明,文中方法在确保良好的检测精度的前提下,可以进一步简化网络的复杂程度、减少计算量,并且能够较好地解决道路交通场景下的遮挡目标和小目标检测的问题。In view of the fact that the object detection algorithm model in road traffic scene takes up more system resources and its detection accuracy for small objects and occluded objects is low,a lightweight object detection method based on improved YOLOv5s is proposed.Some modules with large computation burden in the backbone network is replaced with Ghost module or deeply separable convolution module,which can reduce network size and improve inference speed.The SE(Squeeze⁃and⁃Excitation)module is added to the backbone network to filter the channel feature information and improve the feature expression ability.Repulsion loss function is taken as bbox loss function,which reduces the distance between the predicted frame of the object and the matched object frame and increases the distance between the predicted frame and the surrounding unmatched object frame.DIoU⁃NMS is used as the post⁃processing method,which is used to not only analyze the overlapping area,but also calculate the center point distance between the two boxes according to the suppression criterion,so as to improve the accuracy of object detection under occlusion.The data sets of traffic participants in the traffic scene is constructed.It contains 61200 images in total,including 48960 as the training sets and 12240 as the test sets.It has been verified on the mainstream hardware platform GTX1080 GPU.The mAP(mean average precision)of the proposed method is 85.83%,its FPS(frames per second)is 76.9 f/s,and its model size is 25.6 MB.The mAP of the proposed method is 0.86%higher than that of YOLOv5s,and its FPS and model size are superior to those algorithms of YOLOv4 and YOLOv5.The experimental results show that the proposed method can further simplify the complexity of the network and reduce the computation burden while ensuring good detection accuracy,and can better solve the problem of occlusion and small object detection in road traffic scenes.

关 键 词：深度学习 轻量级卷积神经网络 目标检测 YOLOv5s算法 Ghost模块 深度可分离卷积 损失函数 遮挡目标 

分 类 号：TN911.73-34[电子电信—通信与信息系统]