Matjaž Jogan (1999) . Prešeren awards for students.

Full text not available from this repository.

Abstract

With the increasing needs for self-localisation of mobile systems in the areas of mobile robots, virtual reality and augmented reality, more and more emphasis is given to the problem of localisation and orientation in space. Particularly interesting are the solutions that use only visual information. Most of this methods are based on a geometric 3D model gained from special depth sensors or from correspondence estimation and tracking over periods of time. Recent biological evidence however suggests that position and orientation can be estimated from an adequately compressed set of environment snapshots and their relationships. Localisation can be then performed as recognition of the momentary snapshot in terms of the sensor position. In our work we studied the possibilities of combining such an approach with the technique of compressing a set of panoramic snapshots in the eigenspace. Because of the special characteristics of cylindric panoramic images in means of their changing when the sensor is being rotated, we propose some rotational-invariant representations which allow us to represent a position in space with just one image or its representation. From the results of localisation in eigenspace we can conclude which are the preferred characteristics for a rotational invariant representation of panoramic images. As an alternative option we also test the localisation based on eigenspaces built from panoramic snapshots in multiple orientations.

Actions (login required)