Locally adaptive isotropic detection of corners in object boundaries

Detection of corners is an important task in computer vision to capture discontinuous boundaries of objects of interest. Present operators designed to detect boundaries having sharp corners often produce unsatisfactory results because the points detected can also be an isolated point, ending of a thin line or a maximum curvature region of a planar curve. A novel corner detection operator, capable of detecting corner points that exist only on the boundary of an object, is presented in this paper. Initially, candidate corner points are detected by exploiting local neighbourhood intensity information and associated connectivity pattern around the center point within a local window. Further verification is done to confirm whether the detected corner point is on the boundary of the targeted object. As the proposed operator is isotropic, it covers all the orientations and corner angles by performing a single computation step within the local window. The performance of the operator is tested with both synthetic and real images and the results are compared with other major corner detectors.