Localización de robots móviles en espacios inteligentes utilizando cámaras externas y marcas naturales

Zusammenfassung

This thesis deals with the problem of mobile robot localization using static cameras placed in the environment. The presented approach is based in the idea of “Intelligent Space”, where a distributed intelligence controls cameras and robots to serve in a certain task. The previous works that shares the same approach are focused on placing artificial landmarks on the robots. This thesis focuses on approaches that do not need previous knowledge about the robot and make use of the natural appearance of the robot. The localization process proposed in this thesis is based on natural landmarks, which are detected in the image plane of the set of cameras and correspond to a 3D model of the robot. The proposed localization system is divided in two steps. Firstly, a initialization step obtains the 3D model of the robot and its initial pose. Secondly, using a sequential approach, the pose of the robot is obtained at each time instant. The initialization step is solved in this thesis for any number of cameras using a structure-from-motion approach, where odometry serves as a metric reference in the single camera case. Besides, the proposed approach avoids using natural landmark correspondences between multiple cameras, which allows to initialize the 3D model for non-overlapping views. The sequential step proposed in this thesis uses the 3D model, obtained in the aforementioned initialization step, for retrieving robot’s pose in a estimation-correction scheme. This step includes robust techniques that remove outliers from the measurements. In addition to the filtering approach, a non-iterative and of low complexity solution to the mPnP (multiple perspective n point) problem is proposed. The probabilistic approach performs coherent data fusion between all information available and it allows to estimate uncertainty in the obtained robot’s pose. The solution presented in this thesis has been experimentally assessed using synthetic generated data and experiments from a real environment with cameras, robots and obstacles. The resulting method is proved to be stable against occlusions and illumination changes, which makes it suitable to real situations.