Qualitative tracking of 3-D objects using active contour networks

Research in 3-D object tracking is typified by approaches which attempt to recover the six degrees of freedom of an object in each frame, e.g., [ l l , 14, 6 , 171. Once a correspondence between image and model features is determined, changes in the positions of image features in successive frames is used to update the pose of the object. Although providing accurate pose of the object at each frame, these techniques require an exact geometric specification of the object. At the other extreme, active contour methods have been used to track objects in the image with little or no knowledge of the 3-D object, e.g., [8, 13, 3, 21. However, these methods can only track the position of the object in the image and not its orientation. In this paper, we track changes in the appearance of the object as it moves from one frame to the next. At a symbolic level, an aspect graph clusters all the views of an object into a set of topologically distinct classes in terms of which surfaces of an object are visible from a given viewpoint (Koenderink and van Doorn [9]). Two nodes (or aspects) in the aspect graph are connected