改进YOLOv7的道路多目标检测算法  

Improved YOLOv7 object detection algorithm for multiple road targets

作　　者：张琦 张赛军[1] 周广生 谢豪 ZHANG Qi;ZHANG Saijun;ZHOU Guangsheng;XIE Hao(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,China)

机构地区：[1]华南理工大学机械与汽车工程学院,广州510640

出　　处：《重庆理工大学学报（自然科学）》2024年第11期27-34,共8页Journal of Chongqing University of Technology：Natural Science

基　　金：广东省重点领域研发项目(2020B010184002)。

摘　　要：道路环境感知是自动驾驶任务中的重要组成部分,为解决道路环境感知中小目标检测困难、检测目标尺寸不一致以及检测目标的遮挡给检测任务带来的困难,提出一种深度学习增强方法以提高目标检测性能。设计了Bottleneck-ELAN(bottleneck-efficient layer aggregation networks)模块作为主干,加强了模型的特征提取能力。使用Gather-and-Distribute(GD)机制实现了特征图之间跨尺度的直接融合,解决了颈部网络的信息丢失问题。此外,采用Complete-IOU(CIOU)和Normalized Wasserstein Distance(NWD)相结合的损失函数组,解决了单一IOU损失函数对不同尺度物体位移敏感性不一致和平滑性差的问题。结果表明,改进后的模型在BDD100K数据集上的平均精度均值达到了43.4%,相较于原始的YOLOv7算法提高了3.1%,并且在小目标检测中精度提升更为明显,达到10%。The perception of road environment is an important component of autonomous driving tasks.To overcome the difficulties in detecting small targets,inconsistent detection target sizes,and occlusion of detection targets in road environment perception,we propose a deep learning enhancement method to improve target detection performance.First,the Bottleneck-ELAN module is developed as the backbone to enhance the model’s feature extraction capability.The Gather-and-Distribute(GD)mechanism are also employed to achieve cross-scale fusion between feature maps,addressing the issue of information loss during feature fusion across different scales.Then,a combination of the Complete-IoU(CIoU)and Normalized Wasserstein Distance(NWD)loss functions is employed to address the inconsistency in sensitivity to object displacement and the smoothness disparity present in the single IoU loss function.Our experiment shows the average accuracy of the improved model on the BDD100K dataset reaches 43.4%,3.1%higher than that of the original YOLOv7 algorithm.Moreover,the accuracy of small object detection improves even more markedly,up by 10%.

关 键 词：计算机视觉 目标检测 深度学习 YOLOv7算法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]