Scale-Space Features in 1D Omnidirectional Images

Abstract

Abstract. We define a family of interest operators for extracting features from one-dimensional omnidirectional images, and explore the utility of such features for navigation and localization of a mobile robot equipped with an omnidirectional camera. A 1D circular image, formed by averaging the scanlines of a cylindrical panorama, provides a compact representation of the robot’s surroundings. Feature detection proceeds by applying local interest operators in the scale space of the image. The work is inspired by the recent success of similar operators developed for 2D images. The advantages of using features in omnidirectional 1D images are fast processing times and low storage requirements, which allows a dense sampling of views. We present experimental results on real images that demonstrate that our features are insensitive to noise, illumination variations, and changes in camera orientation. We also demonstrate that most features remain stable over changes in viewpoint and in the presence of some occlusion, thus allowing reliable tracking of features through sequences of frames. 1