Abstract |
A very promising field in the domain of machine vision applications is the navigation of autonomous mechanical systems. These systems move in various envirnments trying to acheve a set of prescribed goals, and their capabilities include egomotion estimation, obstacle avoidance, indepentent motion detection, motion planning and space mapping. In particular, egomotion estimation constitutes one of the most principal tasks of an autonomous system, since the continuous control of its own motion, not only is used for self-reference with respect to the environment, but also is indispensable for making easier the composition of higher level functions. The contribution of this work consists in the implementation and study of four algorithms which compute approximately egomotion parameters from successive images, which the moving observer-dimension motion, six egomotion parameters, three of which represent the translation of the autonomous system, while the rest three represent its rotation. The approach followed is based on extracting information from the normal flow field, which actually represents the displacement of the edges in the scene, with respect to the direction perpendicular to the edges. Despite the fact that the motion information contained in the normal flow field is not complete, the normal flow field is acheived by much easier methods. The sign alternations of the translative and rotative components of the normal flow vectors yield a set geometrical constraints to the position on the image of two special points, FOE and AOR, which characterize the movement parameters, and are charaacterized by them. Particularly, FOE and AOR are intersections of the image plane with the vectors which represent respectively the translation and the rotation of the observer. The goal of the designed algorithms is to specify small areas in the image that contain FOE and AOR. Exploiting the geometrical properties of normal flow consists in specifing simple criteria which test the appearance of specific patterns in the image. The simplicity of these criteria is of particular importance, since it allows fast computation, crtical to the rapid response of the system. The patterns-search criteria are applied to three groups of normal flow vectors, namely vectors of constantdirection and constant norm, as well as vectors belonging to the so-called coaxis and copoint sets, constituting two particular classes of plane vectors. The performance of the algorithms has been tested with synthetic - mainly - as well as real data. The regions where FOE and AOR belong have ben successfully discovered. the size of these regions was small enough in the case of synthetic images, however, it appeared larger with real ones. This increase is mainly due to the diminution of available useful information extracted from the normal-flow field.
|