Adaptive Dominant Points Detector for Visual Landmarks Description

This paper describes a visual landmark detection and description system for mobile robotic environment mapping. At the detection stage, the proposal employs a perceptual-based hierarchical mechanism which has been previously used to successfully solve the segmentation of natural images. At the description stage, visual landmarks are characterized by a kernel-based representation and by the set of dominant points extracted from their outer boundaries. These dominant points allow to represent the landmark region by a set of stable points which can be located –by its position and uncertainty– on the 3D real world. Although all system modules are considered and evaluated in this paper, it is specifically focused on the dominant points detection. Hence, this paper evaluates the performance of a fast curvature–based approach for adaptive detection and its robustness to shape affine transformations.