Curve contour detection based on Helmholtz principle and application to road white line detection
Curve contour detection based on Helmholtz principle and application to road white line detection
复制标题
基于亥姆霍兹原理的曲线轮廓检测及其在道路白线检测中的应用
DOI：
10.1299/transjsme.17-00203
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
H. Mouri
中科院分区：
文献类型：
--
作者：
Hiroyuki Furushou;H. Mouri
提出了一种基于亥姆霍兹原理的道路白色线曲线轮廓检测算法。白色线检测广泛应用于主要用于高速公路或主要干线道路的驾驶员支持系统中。由于普通道路将是自主车辆的作业区域目标，因此有必要开发一种新的检测算法，可以处理各种类型的道路。通过分析Hough变换的局限性，提出了一种基于亥姆霍兹原理的边缘特征的无模型算法。该特征与Hough定义的直线边数特征基本相同，但有两个显著点。一是对计数区域的限制，二是既能检测直线又能检测曲线的计数方法。文中介绍了卷积神经网络的实现方法，并讨论了可调参数与检测性能和处理时间的关系。通过对测试图像和车载摄像机拍摄的图像进行对比，说明了Hough变换和机器学习算法BEL等传统轮廓检测方法的优越性。我们证明了所提出的算法，可以适用于不同的道路环境，但几乎不受噪声的影响，可以实现。
We propose curve contour detection algorithm for road white line detection based on Helmholtz principle. White line detection is widely used in driver support systems used mainly in highway or major arterial road. As the common road will be the target of operational area for autonomous vehicle, it is thought to be necessary to develop a new detection algorithm that can deal with various types of road. This paper proposes model-less algorithm that is constructed on a new edge feature inspired by Helmholtz principle through the analysis of the limit of Hough transform. This feature is basically same as Hough defined feature of edge count on the line except two remarkable points. The one is the restriction of count area and the other is the way of count which affords to detect curve line as well as straight line. Implementation by convolutional neural network is explained and the relation between tunable parameters and the detection performance as well as the processing time are discussed. Comparison between conventional methods such as Hough transform or machine learned contour detection algorithm BEL is explained for test image and images taken by on-board camera to show the superiority of proposed algorithm. We demonstrate that proposed algorithm that can apply to diverse road environments but is hardly affected by noise can be realized.