基于轮廓局部线性拟合度估计的角点检测方法

Corner Detection Method Based on Local Linear Fitting Estimation of Contours

角点检测在计算机视觉、模式识别等领域有着广泛的应用,而基于轮廓的角点检测方法是重要的一个分支,但目前此类算法还存在对轮廓噪声和局部变化敏感,定位性能较差等不足。为了克服这些缺点,提出了一种新的基于图像轮廓局部线性拟合误差的角点检测算法。具体步骤为:首先定义一条带参直线,然后利用最小二乘拟合技术估计参数,使得直线与局部离散点集的拟合误差最小,最后将最优参数下的最小拟合误差作为曲线的局部弯曲程度估计。与现有基于曲率的角点检测算法相比,提出的算法使用距离而不是导数来定义角点响应函数,对局部变化不敏感;此外采用较大的支持域和正则化方案,进一步增强了对噪声和仿射变换的鲁棒性。实验结果表明,在平均重复率(Average Repeatability)和精确度(Accuracy)的评价标准中,提出的算法在高斯白噪声和仿射变换下均有优异的表现。

Corner detection has extensive applications on computer vision and pattern recognition,among which contour-based corner detection methods are an import branch.However,contour-based corner detectors suffer from some weaknesses,such as,sensitive to noises and local variations of boundaries and performing poorly on localization et.al.To overcome the aforementioned drawbacks,this paper proposes a new contour-based corner detection method using local linear fitting error technique.In detail,a parameterized line was defined first;then least square fitting technology was employed to minimize the fitting error between the line and the local discrete point set on a contour;finally the fitting error was considered as the estimation of the bending degree of the contour.Compared with the existing curvature-based corner detection methods,the proposed method uses Euclidean distance instead of derivative for constructing the corner response function,which is insensitive to local variations.In addition,the robustness to noise and Affine transformation is further enhanced by using a larger support region and a regularization scheme.Experimental results show the superiority of the proposed method in terms of the average repeatability,accuracy and localization error under Gaussian white noise and affine transformations.