高级检索
    李旭东, 张建明, 谢志鹏, 王进. 基于三尺度嵌套残差结构的交通标志快速检测算法[J]. 计算机研究与发展, 2020, 57(5): 1022-1036. DOI: 10.7544/issn1000-1239.2020.20190445
    引用本文: 李旭东, 张建明, 谢志鹏, 王进. 基于三尺度嵌套残差结构的交通标志快速检测算法[J]. 计算机研究与发展, 2020, 57(5): 1022-1036. DOI: 10.7544/issn1000-1239.2020.20190445
    Li Xudong, Zhang Jianming, Xie Zhipeng, Wang Jin. A Fast Traffic Sign Detection Algorithm Based on Three-Scale Nested Residual Structures[J]. Journal of Computer Research and Development, 2020, 57(5): 1022-1036. DOI: 10.7544/issn1000-1239.2020.20190445
    Citation: Li Xudong, Zhang Jianming, Xie Zhipeng, Wang Jin. A Fast Traffic Sign Detection Algorithm Based on Three-Scale Nested Residual Structures[J]. Journal of Computer Research and Development, 2020, 57(5): 1022-1036. DOI: 10.7544/issn1000-1239.2020.20190445

    基于三尺度嵌套残差结构的交通标志快速检测算法

    A Fast Traffic Sign Detection Algorithm Based on Three-Scale Nested Residual Structures

    • 摘要: 智能驾驶对交通标志自动检测的实时性及鲁棒性有着极高要求.目标检测中YOLOv3-tiny检测算法是轻量网络,实时性好、但准确率不高.将YOLOv3-tiny检测算法作为基础网络,提出了一种三尺度嵌套残差结构的交通标志快速检测算法.首先,在基础网络上采用逐像素相加的跨层连接,并未增加特征图的通道数,同时网络中形成1个小残差结构.其次,通过同样的跨层连接方式,增加了1层空间分辨率更高的预测输出,使得该尺度输出包含更丰富的空间信息,进而构成大残差结构.最终,将2个残差结构进行嵌套,形成了1个三尺度预测的嵌套残差网络模型,使得Tiny检测算法的部分主网络位于这2个残差结构中,起到3次调参的作用.实验结果表明:提出的算法能够快速鲁棒地检测真实场景中的交通标志.在德国交通标志检测数据集(German traffic sign detection benchmark, GTSDB)上交通标志总F\-1值为91.77%、检测时间为5ms;在长沙理工大学中国交通标志检测数据集(CSUST Chinese traffic sign detection benchmark, CCTSDB)上指示、禁令、警告三大类交通标志F\-1值分别为92.41%,93.91%,92.03%,检测时间为5ms.

       

      Abstract: Automatic driving technology has high requirements for real-time and robustness of traffic sign detection in real world. The YOLOv3-tiny model is a lightweight network with good real-time performance in the object detection, but its accuracy is not high. In this paper, we use YOLOv3-tiny as the basic network and propose a fast traffic sign detection algorithm with three-scale nested residual structure. Firstly, shortcut based on pixel by pixel addition is employed in the YOLOv3-tiny network. It does not increase the number of feature map channels, and a small residual structure is formed in the network at the same time. Secondly, the predictive output with higher spatial resolution is also added through the shortcut, which contains more abundant spatial information, thus forming a large residual structure. Finally, the two residual structures are nested to form a three-scale predictive nested residual network, which makes the main network of Tiny located in these two residual structures and the parameters can be adjusted three times. The results show that the proposed algorithm can quickly and robustly detect traffic signs in real scenes. The F\-1 value of total traffic signs achieves 91.77% on German traffic sign detection benchmark and the detection time is 5ms. On CSUST Chinese traffic sign detection benchmark, F\-1 values of the Mandatory, the Prohibitory and the Warning are 92.41%, 93.91% and 92.03% respectively, and the detection time is 5ms.

       

    /

    返回文章
    返回