Study of improving the stability of SUSAN corner detection algorithm

Abstract

The corner detection has become an essential and fundamental procedure in many computer vision problems, such as image registration, image matching, scene analysis, motion and structure from motion analysis, object recognition, etc. Smallest Univalue Segment Assimilating Nucleus (SUSAN) is one of the most excellent methods which are robust to noise and less affected by rotation. However, it could not detect all the true corners and generate some false corners in some special case. To solve these problems, an improved SUSAN corner detector is proposed and its performance is compared with SUSAN corner detection. With the improved SUSAN, a corner point is judged based on gray level values of the pixels in a circular neighborhood of the nucleus which is the same as the conventional SUSAN, however, the improved SUSAN calculates the number of the pixels in the univalue adjoining nucleus and connected segment rather than calculate the number of the pixels of univalue nucleus in the neighborhood. Due to this improvement, the improved SUSAN can not only inherit the main merits but also avoid the fatal fault of conventional SUSAN. Experimental results have demonstrated that the improved SUSAN corner detection is accurate and efficient.