交通运输系统工程与信息

交通运输系统工程与信息 ›› 2019, Vol. 19 ›› Issue (3): 48-53.

• 智能交通系统与信息技术 • 上一篇    下一篇

基于感兴趣区域的CNN-Squeeze 交通标志图像识别

张秀玲*a, b，张逞逞 a，周凯旋 a   

  1. 燕山大学 a. 河北省工业计算机控制工程重点实验室； b. 国家冷轧板带装备及工艺工程技术研究中心，河北 秦皇岛 066004
  • 收稿日期:2018-11-15 修回日期:2019-01-16 出版日期:2019-06-25 发布日期:2019-06-25
  • 作者简介:张秀玲(1968-)，女，山东章丘人，教授，博士.
  • 基金资助:

    河北省自然科学基金/ Natural Science Foundation of Hebei Province(E2015203354)；河北省教育厅科学研究计划河北省高等学校自然科学研究重点项目/ Science and Technology Research Key Project of High School of Hebei Province (ZD2016100)；2016年燕山大学基础研究专项(理工类)培育课题/Basic Research Special Breeding Project Supported by Yanshan University (16LGY015).

Traffic Sign Image Recognition via CNN-Squeeze Based on Region of Interest

ZHANG Xiu-linga, b, ZHANG Cheng-chenga, ZHOU Kai-xuana   

  1. a. Key Laboratory of Industrial Computer Control Engineering of Hebei Province; b. National Engineering Research Center for Equipment and Technology of Cold Strip Rolling, Yanshan University, Qinhuangdao 066004, Hebei, China
  • Received:2018-11-15 Revised:2019-01-16 Online:2019-06-25 Published:2019-06-25

PDF

摘要：

在公路交通中，针对复杂环境下交通标志识别率不高的问题，提出了一种基于 Kmeans对图像聚类，切割图像感兴趣区域(Regions of Interest, ROI)，并利用方向梯度直方图特征(Histogram of Oriented Gradient, HOG)与卷积运算，特征加权(CNN-Squeeze)相结合的交通标志识别方法.首先，采用 K-means对交通标志图像进行三角形、圆形图像二聚类，并利用制作的切割模板切割 ROI 并提取 HOG 特征；然后，利用卷积神经网络 (Convolutional Neural Network, CNN)对 HOG特征进行过滤、降维，并通过 Squeeze网络对过滤后的二次特征进行重要性标定；最后，训练该网络模型并实现对交通标志的识别.仿真结果表明，与 BP网络、SVM 及CNN对比，本文方法在保证训练时间的同时，识别精度达到98.58%.

关键词: 智能交通, K-means, 感兴趣区域, CNN-Squeeze, 交通标志识别

Abstract:

In highway traffic, in view of the low recognition rate of traffic signs in complex environments, a traffic sign recognition method based on K-means image clustering and image- cutting ROI is proposed, which combines HOG feature with convolution operation feature weighting (CNN-Squeeze). Firstly, K-means is used to perform clustering of triangles and circular images on the traffic sign image, and the ROI is extracted by using the cutting template. Then, the HOG feature is filtered and reduced by the CNN network, and the filtered secondary features are calibrated by the Squeeze network. Finally, the traffic sign is recognized by using the trained neural network. Simulation results show that compared with BP network, SVM and CNN network, this method can guarantee the training time and the recognition accuracy reaches 98.58%.

Key words: intelligent transportation, K-means, regions of interest, CNN-Squeeze, traffic sign recognition

中图分类号: