摘要
针对YOLOv5算法在检测交通信号灯过程中存在的误检、漏检及模型特征提取能力不足等问题,提出改进的交通信号灯检测算法ST-YOLOv5。首先,去除主干网络末端的卷积层,并在末端加入基于窗口和移动窗口的多头注意力机制;其次,设计由高分辨率、低水平特征图组成的检测层;最后,通过增加浅层高分辨率特征与深层强语义特征之间的跨层级连接来缓解由于通道数减少而造成的小目标信息丢失问题,并在多层特征聚合后加入有关通道和位置的注意力机制。实验结果表明,在BDD100K数据集中,此改进算法对交通信号灯的检测精度达到70.10%,有效减少了误检、漏检等问题。
Aiming at the problems of false detection,missed detection,and insufficient model feature extraction ability in YOLOv5 traffic lights detection,an improved YOLOv5 traffic light detection algorithm ST-YOLOv5 is proposed.Firstly,the convolutional layer at the end of the backbone network is removed and a multi head attention mechanism based on window and moving window at the end is added.Secondly,a detection layer composed of high-resolution and low-level feature maps is designed to specifically detect small targets.Finally,by increasing the cross-level connection between low-level high-resolution features and high-level high-semantic features,the problem of small target information loss caused by the reduction of channel numbers is alleviated,and attention mechanisms for channels and positions are added after multi-layer feature aggregation.Experimental results show that on the BDD100K dataset,ST-YOLOv5 has a detection accuracy of 70.1%for traffic signal lights,effectively reducing false detections and missed detections.
作者
雷亮
秦兰瑶
张文萍
和圆圆
梁明辉
尹衍伟
陈毅
LEI Liang;QIN Lanyao;ZHANG Wenping;HE Yuanyuan;LIANG Minghui;YIN Yanwei;CHEN Yi(School of Intelligence and Technical Engineering,Chongqing University of Science and Technology,Chongqing 401331,China)
出处
《重庆科技学院学报(自然科学版)》
CAS
2023年第3期94-103,共10页
Journal of Chongqing University of Science and Technology:Natural Sciences Edition
基金
2021年重庆市属本科高校与中科院所属院所合作项目“工业互联网内生产安全关键技术研究与协同创新”(HZ2021015)。