Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (7): 164-170.DOI: 10.3778/j.issn.1002-8331.2007-0117

Previous Articles     Next Articles

YOLOv3 Traffic sign Detection based on SPP and Improved FPN

LIU Ziyan, YUAN Lei, ZHU Mingcheng, MA Shanshan, CHEN Linzhouting   

  1. 1.College of Big Data and Information Engineering, Guizhou University, Guiyang 550025, China
    2.School of Aerospace Engineering, Guizhou Institute of Technology, Guiyang 550003, China
  • Online:2021-04-01 Published:2021-04-02

融合SPP和改进FPN的YOLOv3交通标志检测

刘紫燕,袁磊,朱明成,马珊珊,陈霖周廷   

  1. 1.贵州大学 大数据与信息工程学院,贵阳 550025
    2.贵州理工学院 航空航天工程学院,贵阳 550003

Abstract:

Aiming at solving the problems of small size, low resolution and insignificant features in traffic sign targets detection, an improved network model of YOLOv3 is proposed. After using the color enhancement method to enhance the traffic sign data, the FPN structure in the original network is improved, retaining the large-scale prediction with a scale of 52×52 in the original network, then it builds a larger-scale prediction with a scale of 108×108 by using the feature map of the second down-sampling output in the YOLOv3 network. In order to solve the problem of image size and distortion, it uses pooling operations with fixed block sizes of 5, 9, and 13 before the detection layer. And then, the separately output features are merged with the original feature map to achieve the same output size for different input sizes. Finally, the K-means clustering algorithm is used to cluster the TT100K traffic sign data set, the initial candidate box of the network is redefined, and the YOLOv3 network model, the improved YOLOv3 network model and other small target detection algorithms are used to compare experiments on the TT100K data set. The results show that the improved YOLOv3 network model can detect traffic signs more effectively, and the average detection accuracy of the improved YOLOv3 network model is 8.3%, 6.1% and 4.3% higher than that of the original YOLOv3 network model at three scales. When the FPS changes little, the recall rate and accuracy are significantly improved. At the same time, the improved YOLOv3 algorithm has better detection accuracy and real-time performance than other small target detection algorithms.

Key words: object detection, traffic sign, YOLOv3, data enhancement, large-scale prediction

摘要:

针对交通标志目标检测尺寸较小、分辨率低、特征不明显问题,提出一种改进的YOLOv3网络模型。在利用颜色增强方法对交通标志进行数据增强后,改进原网络中的FPN结构,保留原网络中52×52的大尺度预测,然后利用YOLOv3网络中第二次下采样输出的特征图建立108×108的更大尺度预测。为了解决图像尺寸和失真的问题,在检测层前使用固定分块大小为5、9、13的池化操作,再将输出的特征与原来的特征图进行融合,从而实现对不同尺寸的输入得到相同大小的输出。最后,利用K-means聚类算法对TT100K交通标志数据集进行聚类分析,重新定义网络的初始候选框,使用YOLOv3网络模型和改进的YOLOv3网络模型以及其他小目标检测算法在TT100K数据集上进行对比实验。实验结果表明,改进后的YOLOv3网络模型能更有效的检测交通标志,其检测的平均精确度在三个尺度下相对原YOLOv3网络模型分别提升8.3%、6.1%、4.3%,在FPS变化不大的情况下,召回率和准确率都有明显提升,同时,改进后的YOLOv3算法相对其他小目标检测算法具有更好的检测精度和实时性。

关键词: 目标检测, 交通标志, YOLOv3, 数据增强, 大尺度预测