摘要
针对目前自动驾驶过程中对交通标志的识别检测速度慢的问题,提出一种改进的YOLOv5s-Ghost网络模型对交通标志进行识别的方法,在3×3运算核Ghost Net模型框架下,通过两个连续的Ghost模块构建的Ghost Bottleneck模块,并代替C3模块中全部的Bottleneck模块,与跨阶段局部网络(cross-stage position network, CSPNet)模块结合生成Ghost Bottleneck CSP模块。通过调整每个模块中加入Ghost Bottleneck模块的数量,对比实验数据得到最佳网络模型。分别用原网络和新网络对TT100K数据集进行训练,对比实验数据表明,YOLOv5s-Ghost模型的检测精度达95.1%,检测速度达到了52.6 FPS,模型大小压缩了69.3%,在保证原检测精度的情况下提高了网络的检测速度。
In view of the problem of slow detection speed and large network model of traffic signs in the process of automatic driving, an improved YOLOv5s-Ghost network model is proposed to identify traffic signs.Under the framework of 3×3 computing core Ghost Net model, the Ghost Bottleneck module is constructed by two consecutive Ghost modules, which replace all the Bottleneck module of C3 modules and combine with cross-stage position network(CSPNet) to generate Ghost Bottleneck CSP module.The best network model is obtained by comparing the experimental data by adjusting the number of Ghost Bottleneck modules added to each module.The original network and the new network are respectively used to train TT100K data set.Compared experimental data show that the detection accuracy of the YOLOv5s-Ghost model is 95.1%,the detection speed reaches 52.6 FPS,and the model size is compressed by 69.3%,which improves the detection speed of the network while ensuring the original detection accuracy.
作者
徐正军
张强
许亮
XU Zhengjun;ZHANG Qiang;XU Liang(Tianjin Key Laboratory for Control Theory&Applications in Complicated Systems,School of Electrical Engineering and Automation,Tianjin University of Technology,Tianjin 300384,China;Beijing Aerospace Propulsion Institute,Beijing 100076,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2023年第1期52-61,共10页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61975151,61308120)资助项目。
