鉴于数字图像中的拐点通常成为重要的信息载体,因此准确、稳定和实时地检测出拐点便成为拐点检测算法面临的主要问题,针对该问题,提出了一种新的二值图像中拐点的实时检测算法。该算法与传统基于边界链码的拐点检测算法不同,其是首先构建像素的k(k>8)邻域,并将图像中物体的边界表示为k邻域链码;然后根据曲率定义的差分形式计算各边界点处的曲率;最后通过检测曲率直方图的局部峰值精确定位出拐点,并利用拐角内部像素的颜色统计信息迅速判断出拐点的凸凹性.为验证该算法的效果,给出了该算法与4种已有算法的对比实验.结果表明,该算法不仅稳定性、准确性较高,而且算法简单,实时性强,并适合于嵌入式计算环境。

Presents a new real time corner detection algorithm. Corners are important information carriers in object recognition. Accurate, stable and fast detecting corners in digital image are common problems facing to corner detectors. Aiming at these problems and different from traditional corner detection algorithms, based on chain code, the algorithm constructs k(k>8) neighborhood chain codes of pixels and uses these chain codes to describe contours. Based on the differential definition of curvature, a curvature function is derived from k neighborhood chain codes. Corners are detected as those contour pixels, whose curvature the is largest in a lobe of contour curvature histogram. Convex and concave corners can be differentiated quickly by checking color attributes of pixels between corner edges. To validate the algorithm, tests comparing the new algorithm to 4 corner detection algorithms are given. The results show the new algorithm is not only accurate and stable, but also simple and fast, which make the algorithm suitable for the embedded computation environment.