Detector de pontos de interesse baseado em características visuais e de profundidade

Abstract

In computer vision systems, its a common practice to represent patches of an image using a set of carefully chosen points, those points are identified based on desired local features that are convenient to the application. Examples of such features are edges, corners, blobs and ridges. This points are commonly denoted as keypoints. One of the first steps in numerous computer vision tasks is the extraction of keypoints. Despite the large number of works proposing image keypoint detectors, only a few methodologies are able to efficiently use both visual and geometrical information. In this work we introduce KVD (Keypoints from Visual and Depth Data), a novel keypoint detector which is scale invariant and combines intensity and geometrical data through low-cost operations and a decision tree. We present results from several experiments that show that our methodology produces the best performing detector when comparing to state-of-the-art methods, with good repeatability scores for rotations, translations and scale changes, as well as robustness to corruptions on either visual or geometric information. When compared to algorithms with equivalent input type, our algorithm yields the best time results.