Vision based localization of humanoid robots by inverse pose-estimation using a small set of fixed landmark features

Abstract

We propose a vision based methodology to solve the robot localization problem by estimating the pose of the robot's camera with respect to a model of the environment represented by a small set of landmark features. The methodology proposed constructs a regression model based on the pinhole camera projection model by combining the information of several observed features. In this regression model the unknowns are the X and Y coordinates of the location vector of the robot with respect to the environment, and the depths of each of the observed features. We use a RoboCup Humanoid KidSize field as a test bed for experimental validation of this methodology.