Object Recognition Using a Tree-Like Procedure Generated from 3-D Model

This paper presents a vision system which automatically generates an object recognition procedure from a 3-D model, and recognizes the object by executing this procedure. The change in object appearances due to the viewpoint is a main issue in 3-D model vision. In this system, the appearances of an object from various view points are described with visible 2-D figures, such as parallel lines and ellipses. These figures are projections of linear or circular visible elements in the model. Then, the appearance descriptions are compared. Similar figures are extracted and merged into a new description, which represents a common and general appearance of the old ones. This process is iterated and ends in a tree-like procedure. The system search the procedure tree and recognizes the object. At a visited node of the tree, figure is estimated by the recognition procedure and looked for in the image. If such figures are detected, a child of the current node is selected for the next visit. Detected figures are used for estimation at the next node. The number of object candidates increases as the number of detected figures increases, and propagated candidates construct a tree. The system searches the candidate tree in addition to the procedure tree. Experimental results show the efficiency of proposed system. Introduction segments under the constraint given by the recognition procedure. Grimson[4] used a similar method for depth images. Though these approaches have improved efficiency, some problems still remains. The features used for recognition are only local ones, such as line segments or plane surfaces. The number of local features increases as the image complexity does. As a result, search space becomes large which prevents efficient recognition. This paper proposes a new approach to a model-based vision. In this system, the appearance of an object is described with global figures, such as ellipses and parallel lines, instead of local ones. Because the number bf global figures is smaller than that of local features, the search space can be kept small.