Feature Extraction and Grouping for Robot Vision Tasks

Exploiting computer vision for performing robotic tasks, like recognizing a given place in the environment or simply computing the relative orientation of the robot with respect to the environment, requires an in depth analysis of the vision modules involved in such computations. In this paper, we have considered three types of computations: local computations (for the estimation of junctions), local-to-global computations (for finding a geometric sketck) and voting-accumulation computations (for obtaining the relative orientation of the robot). We have addressed the analysis of the latter modules from the point of view of three practical requirements: reliability (robustness), efficiency and flexibility. These requirements are partially fullfiled by the methodology used: the three modules (junction detection, grouping, and orientation estimation) share elements of Bayesian inference. Sometimes, as in the case of junction detection, these elements yield statistical robustness. In other cases, as in the grouping module, the Bayesian formulation has a deep impact in the reduction of computational complexity. The Bayesian integration of feature extraction, grouping and orientation estimation is a good example of how to get flexibility by exploiting both the visual cues and the prior assumptions about the environment in order to solve a given task. On the other hand, as the basic visual cues are edges and there are fundamental limits regarding whether certain tasks relying on edge cues may be solved or not, independently of the algorithm, we also stress the convenience of having this bounds in mind in order to devise practical solutions for robotics tasks driven by computer vision. 103.