Feature Extraction and Grouping for Robot Vision Tasks
Miguel Cazorla, Francisco Javier Escolano · Pro Literatur Verlag, Germany eBooks · 2005
IntroductionThis paper focuses on feature extraction and perceptual grouping in computer vision.Our main purpose has been to formulate and test new methods for feature extraction (mainly those related to corner identification and junction classification) and grouping (through junction connection).Our purpose is to build a geometric sketch which can be useful for a wide range of visual tasks.The context of application of these methods, robot vision, imposes special real-time constraints over their practical use, and thus, the following requirements are observed: efficiency, robustness, and flexibility.The paper is structured in three parts which cover junction classification, grouping, and estimation of the relative orientation of the robot with respect to the environment (an example of visual task).We follow a bottom-up exposition of the topics (Cazorla, 2000):• Junction Classification: It relies on combining corner detectors and template matching.Corner detection is performed through two well known operators: SUSAN and Nitzberg.These operators provide an initial localization of the junction center.Given this localization, junction classification is posed in terms of finding the set of angular sections, or wedges, that explain as better as possible the underlying evidence in the image.We propose two greedy methods which rely on elements of Bayesian inference.These methods are tested with indoor and outdoor images in order to identify their robustness to noise and also to bad localizations of junction centers.Experimental results show that these methods are saver than other methods recently proposed, like Kona, and their computational efficiency is even better.However, we have detected some problems due to the local extent of junction detection.These problems and those derived from bad center localization can be aliviated through local-to-global interactions, and these interactions are inferred by grouping processes.• Grouping: Using classified junctions as starting elements, this stage is performed by finding connecting paths between wedge limits belonging to pairs of junctions, and these paths exist when there is sufficient contrast or edge support below them.Given that corners are usually associated to points of high curvature in the image, it can be assumed that connecting paths must be smooth if they exist.Contrast support and smoothness are quantified by a cost function.As it can be assumed that there will be no more than one path between two junctions through a pair of wedge limits, such a path can be found by the Bayesian version of the well known A* algorithm.This method recently proposed searches the true path, according to the cost function, in a population of false path, instead of the best path among a population of possible paths.