Abstract
Corner detection is a main concern in many computer vision applications like object recognition or image matching. Furthermore, detection is usually performed over the contour of the objects. This paper presents a novel algorithm to detect corners based on CSS (curvature scale space). A multi-scale curvature polynomial is defined as the sum or product of the curvature under all scales of the contour. The new method can not only enhance curvature extreme peaks effectively, but also suppress noise and trivial details and prevent smoothing some corners with the augment of the scale. On the other hand, the detected corners belonging to the concave or convex can be judged by the result sign of the curvature polynomial. Experiment results show that the new method is more effective in corner detection than the other algorithms mentioned in the paper.