Topological and Metric Robot Localization through Computer Vision Techniques

Abstract

Nowadays, robotic applications based on vision sensors have become widespread, but there is still a gap between these applications and the pure computer vision developments. Sometimes this separation can be due to the lack of communication between both research communities or to the divergence in their objectives. Other times this difference is due to the inadequacy of the methods for certain tasks, e.g., there are computer vision methods which can not be applied for robotic tasks due to their computational complexity. However, this can be solved many times just with a slight adaptation of the techniques. Many works during the last years have developed vision based methods for robotic tasks such as control [5], automatic topological map building [23], topological localization [10], or Simultaneous Localization and Mapping [3]. This work is focused on the application of computer vision techniques for robot global selflocalization, a fundamental issue for any autonomous device. Both topological and metric localization are taken into account, as the two of them have huge similarities with computer vision applications. On the one hand, topological localization usually consists of identifying the current location of our mobile device in a higher cognitive level than just metric units, for example identifying the room where the robot currently is. This could also be named room/scene identification. Object recognition is an important issue in computer vision research, with many works and important results in the previous years, e.g. [11], [8] or [19], that could be adapted for scene recognition. For instance, in [12] a room identification technique was presented, that mixes range and camera information and is based on a learning method typically used for object classification/recognition (AdaBoost). On the other hand, the metric localization as well as the Simultaneous Localization and Mapping (SLAM)