中文核心期刊
CSCD来源期刊
中国科技核心期刊
RCCSE中国核心学术期刊

重庆交通大学学报(自然科学版) ›› 2021, Vol. 40 ›› Issue (10): 178-184.DOI: 10.3969/j.issn.1674-0696.2021.10.21

• 交通+大数据人工智能 • 上一篇    

基于改进VGG模型的低照度道路交通标志识别

赵树恩,刘伟   

  1. (重庆交通大学 机电与车辆工程学院,重庆 400074)
  • 收稿日期:2021-02-10 修回日期:2021-06-12 出版日期:2021-10-20 发布日期:2021-10-29
  • 作者简介:赵树恩(1972—),男,陕西洋县人,教授,博士,主要从事智能汽车与自动驾驶方面的研究。E-mail:zse0916@163.com 通信作者:刘伟(1996—),男,重庆合川人,硕士研究生,主要从事车路协同方面的研究。E-mail:1547233037@qq.com
  • 基金资助:
    国家自然科学基金项目(52072054);重庆市基础研究与前沿探索项目(cstc2018jcyjAX0422)

Low-Illumination Road Traffic Sign Recognition Based on Improved VGG Model

ZHAO Shuen, LIU Wei   

  1. (School of Mechatronics and Vehicle Engineering, Chongqing Jiaotong University, Chongqing 400074, China)
  • Received:2021-02-10 Revised:2021-06-12 Online:2021-10-20 Published:2021-10-29

摘要: 针对低照度情况下道路交通标志图像亮度偏低、饱和度过高、图像模糊、识别不精确等问题,提出一种基于膨胀卷积-VGG(dilated convolution-VGG,DC-VGG)模型的道路交通标志快速识别方法。首先,运用限制对比度直方图均衡算法(contrast limited adaptive histogram equalization,CLAHE)对图像H(色相)S(饱和度)V(色明度)空间中的V通道均衡化,实现低照度图像亮度增强;其次,在HSV空间中设定阈值分割出指定色彩,通过轮廓检测定位交通标志;然后,基于深度卷积对抗神经网络(deep convolutional generative adversarial networks,DCGAN)对真实的交通标志图像进行数据样本增强,以提高分类模型的鲁棒性;最后,提出DC-VGG轻量化模型实现交通标志快速识别。经验证,该方法达到94.12%的识别准确率,且能在硬件不佳的条件下实时检测。

关键词: 交通运输工程, 交通标志识别, 低照度, DC-VGG, 对比度受限直方图均衡, DCGAN

Abstract: Aiming at the problems of low brightness, high saturation, blurred image, and inaccurate recognition of road traffic sign images under low illumination conditions, a fast road traffic sign recognition method based on the dilated convolution-VGG (DC-VGG) model was proposed. Firstly, the contrast limited adaptive histogram equalization (CLAHE) algorithm was used to equalize the V channel in the image H (hue) S (saturation) V (color lightness) space, so as to achieve low illumination image brightness enhancement. Secondly, a threshold was set in the HSV space to segment the specified colors, and the traffic signs were located through contour detection. Then, based on the deep convolutional generative adversarial networks (DCGAN), the data samples of the real traffic sign images were enhanced, to improve the robustness of the classification model. Finally, the proposed DC-VGG lightweight model was used to realize the rapid recognition of traffic signs. It is verified that the recognition accuracy of the proposed method is 94.12%, and it can detect in real time under poor hardware conditions.

Key words: traffic and transportation engineering, traffic sign recognition, low illumination, DC-VGG, contrast limited histogram equalization, DCGAN

中图分类号: