Fast feature extraction using approximations to derivatives with summed-area images

Accurate and stable identification of feature points is a requirement for such varied applications as wide-baseline stereo, object recognition and simultaneous localisation and mapping. Although a wide variety of feature extraction methods exist, certain aspects remain active areas of research. In this paper, a feature model is proposed which makes use of the summed area images in achieving scale invariance at the loss of theoretical rotational invariance. By making use of approximations to first and second derivatives, as well as the Laplacian, a wide variety of features may be obtained. Additionally, the stability of this method is increased by an improved approach to ordering of features. Evaluation is performed versus other common approaches using tests on precision, recall and information content of the extracted points.