基于改进单级特征图方法的交通标志检测  

Traffic sign detection based on improved one-level feature image method

作　　者：梁天骄 鲍泓 潘卫国 潘峰 胡正坤[3] LIANG Tianjiao;BAO Hong;PAN Weiguo;PAN Feng;HU Zhengkun(Beijing Key Laboratory of Information Service Engineering,Beijing Union University,Beijing 100101,China;College of Robotics,Beijing Union University,Beijing 100027,China;College of Applied Science and Technology,Beijing Union University,Beijing 100012,China)

机构地区：[1]北京联合大学北京市信息服务工程重点实验室,北京100101 [2]北京联合大学机器人学院,北京100027 [3]北京联合大学应用科技学院,北京100012

出　　处：《传感器与微系统》2022年第2期123-127,131,共6页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61802019,61932012,61871039);北京市教育委员会科技计划资助项目(KM201911417003,KM201911417009,KM201911417001);北京联合大学研究生科研创新资助项目(YZ2020K001,YZ2021K001)。

摘　　要：基于深度学习的交通标志牌检测算法取得了突破性的进展,但在检测精度和速度方面仍得不到兼顾。针对此问题,本文在YOLOF算法的基础上提出了一种改进的算法,在YOLOF网络检测分支中融入注意力机制以增强网络对交通标志牌目标的表示,并利用CIoU改进损失函数;使用数据增广模拟自动驾驶过程中的复杂环境,增强检测模型的鲁棒性。对比实验结果表明:本文提出的改进方法具有更高的检测精度,能够达到检测精度和速度的平衡。The deep learning based traffic sign detection algorithm has made breakthrough progress, but the precision and speed of target detection still cannot be balanced.To solve this problem, propose an improved algorithm on the basis of YOLOF algorithm.The algorithm fuses attention mechanism to strengthen representation of network on traffic sign in YOLOF network detection branch, and CIoU is used to improve the loss function.And uses data augment methods to simulate complex environment of automatic driving to enhance robustness of the detection model.The experimental results show that the proposed improved method has higher detection precision, and can achieve balance of detection precision and speed.

关 键 词：目标检测 注意力机制 交通标志牌 单级特征图 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]