The use of quadtrees in a recognition and localization problem

A method for planar object recognition and localization is described. In the learning phase, an image of the object in a reference position is stored as a model. The recognition consists in computing the reference position of the shape and then the corresponding image using a rotation and translation algorithm. The distance used in the recognition process is the area of the exclusive or between the new image and the model. The authors show how to use quadtrees to make this scheme computationally efficient. The method is then extended to the recognition of nonconvex partially occluded objects by extracting concavities and trying to match them with the concavities of the models.<>