Recognition of 3-D objects on complex backgrounds using model based vision and range images

摘要

One of the active research fields in computer vision is the recognition of complex 3D objects. The task of object recognition is tightly bound to background understanding or suppression. Current literature describes the top down approaches as promising but not complete and the bottom-up approaches as not robust. The paper describes a model based vision system in which a commercial 3D computer graphics system has been used for object modeling and visual clue generation. Given the computer generated model image, a conventional CCD camera image and the corresponding scanned 3D dense range map of the real scene, the object can be located in it. The paper deals with how this is done using newly developed segmentation algorithms extracting "focus features" from range images (depth map) of the scene. The system uses the image pyramid of resolution and prediction-verification process. First the authors generate a hypothesis in a low resolution description, giving rough clues for the object boundaries, position and orientation. These regions of interest are then used as the field of comparison with higher resolution models. Such an iterative process is repeated until a given threshold of similarity is reached. Next an intensity image of the model in the scene is created using the available a priori knowledge. Direct correlation is then performed between the model and the "focus feature" of the scene. Illustrative examples of object recognition in simple and complex scenes are presented.