Corner detection in planar curves is related to corner detection in grayscale images which is not addressed here. Characteristic contour points have traditionally been in the focus of the scale space theory [5] that allows for a `natural', although sophisticated and computationally demanding, definition of such points at varying scale. However, in many online applications, especially in industry, processing time is a crucial issue. Computational load is to be minimized without significant loss of robustness.
Various less complicated corner detection algorithms have been developed. A number of frequently cited approaches are discussed in the survey by Liu and Srinath [4], where comparative experimental results are also given. Four of the algorithms tested in [4] are used in our tests as well, namely those by Rosenfeld and Johnston [6], Rosenfeld and Weszka[7], Freeman and Davis [3], and Beus and Tiu [2]. In this study, these algorithms are referred to as RJ73, RW75, FD77, and BT87, respectively. RW75 is a modification of RJ73, while BT87 is a modification of FD77. A summary of the four algorithms is given in section 2.
Although the notion of corner seems to be intuitively clear, no generally accepted mathematical definition exists, at least for digital curves. In a sense, different approaches give different -- but conceptually related -- computational definitions to a visual phenomenon. The lack of unequivocal ground truth makes comparative performance evaluation tests less significant than they could, and should, be.
Here we present a new, fast and efficient algorithm for detection of high curvature points. (For simplicity, they will be called `corner points'.) The parameters of the algorithm are easy to understand and tune to particular sharpness and scale. The new algorithm, referred to as IPAN99, is described in section 3. (IPAN stands for Image and Pattern Analysis group.) Experimental results shown in section 4 compare the new method to the alternative ones mentioned above.