Method for detecting interest points in images using angular signatures

We present an innovative method for detecting interest points (IPs) in grayscale and color images. It is based on the use of angular signatures (ASs), produced by spinning, at each pixel in the image, an "x-tapered, y-derivative, half-Gaussian kernel" in discrete angular steps. By exploiting the AS(s) produced at each pixel, it automatically "classifies" the pixel as being an IP or not. We present preliminary results on synthetic grayscale and real color 2D images, and these confirm the potential value of the method. It can easily be extended from grayscale and color images to images with any number of components, as well as to 3D volumetric images and images on grids of higher dimensionality. It is useful for stereo matching and video tracking.