Topological and Metric Robot Localization through Computer Vision Techniques

This paper explains a vision-based method to obtain both topological and metric localization through a hierarchical process, presented in our previous work [17]. There, global localization is obtained with respect to a visual memory (a topological map built with sorted reference images). The global localization, sometimes known as the ”kidnapped robot problem”, intends to localize the robot only with the current acquisition of the sensors, without any knowledge of previous measurements, oppositely to the continuous localization tasks. The aforementioned localization hierarchy consists of an initial less accurate localization result, in terms of topological information (room identification), which applies object recognition techniques. The second localization result of the hierarchy is a more accurate metric localization. It is obtained through a SFM algorithm for 1D bearing only data [1], [4] based on the 1D trifocal tensor [6]. This kind of data is intuitively extracted from images, Fig. 8.1 shows two examples of 1D bearing only data. On the left, the orientation of point features in omnidirectional images, that is the more stable cue in that kind of images; on the right, another situation where using only 1D is convenient, the horizontal coordinate of vertical lines in conventional images, as these line segments usually have a clear orientation (x-coordinate) but they do not have accurate tips (y-coordinate).

The outline of this paper is as follows. Next section 8.2 is divided in two parts: subsection 8.2.1 details the process used to perform the room/scene recognition and subsection 8.2.2 explains the 1D trifocal tensor estimation and its SFM algorithms. In section 8.3, a brief description of the features that have been studied in our examples is given, followed by section 8.4 with several examples of localization results obtained applying the explained techniques. Finally section 8.5 concludes the work.