一维概率Hough变换的实时鲁棒多圆检测方法

Robust Real-Time Multi-Circle Detection Algorithm Based on 1D Probabilistic Hough Transform

针对现有圆检测算法以像素为投票主体在二维或三维空间进行参数累积,运算复杂度高,难以达到复杂自然场景中的实时检测要求的问题,提出一种以线段为投票主体并基于一维概率Hough变换的实时圆检测算法.首先基于梯度方向对Canny边缘像素实施逐段分割,从中选取满足一定曲率条件的线段作为种子,对每个种子沿半径进行一维概率Hough累积;然后根据峰值大小和位置提取有效圆及其初始半径;最后利用圆的直接最小二乘拟合进一步定位圆半径和圆心.对复杂的自然场景图像进行实验的结果表明,通过选取合适的分割阈值,该算法在确保圆检测可靠性的前提下在速度上显著优于现有算法.

The state of the art in circle detection usually resorts to edge pixels as the voting components to perform parametric accumulation in 2D or 3D space, which generally incurs high computational cost and is thus unable to meet the real-time processing requirements in complex natural scene processing. Using edge sections as voting components, this paper presents a robust real-time circle detection algorithm based on 1D probabilistic Hough Transform. The algorithm first segments Canny edges based on their gradient directions into arc sections, from which seed sections meeting certain curvature criteria are selected. For each seed, a probability-weighted 1D Hough accumulation is then built along the radius dimension to detect a valid circle related to the seed and estimate the initial radius of the circle based on the peak magnitude and peak position of the 1D accumulation. Finally direct circular least square fitting is employed to further pinpoint the radius and center information for the detected circle. The experiment shows, when appropriate segmentation thre-sholds are chosen, the algorithm significantly outperforms the state of the art in processing speed while maintaining high reliability as far as the circle detection in complex natural scene images is concerned.