Abstract
The most important application area of visual pattern recognition is design and development of a computer visions system wheather for robotics or for industrial inspection, capable of recognising and determining the portion of an object in a scene, particularly so when the objects are occluded. In this paper a new computational approach along with the computational results are presented based on the concept of differential geometry for recognition and position determination of partially occluded 2D rigid object and successful extension and implementation of the method for the 3D objects.
For a partially occluded 2D objects in a scene a set of invarient local features are generated initially. Next, based upon matching of local features of the objects in a scene and the models which are considered as cognitive data base, a computer vision scheme is described using AI concept of hypothesis generation and verification of features for the best possible recognition. The method is successfully extended to the real life task of recognition and position determination of partially occluded 3D objects in a scene. The 3D surface informations which may be planer or curved are captured through depth map from range data. For recognition the principal curvatures, mean curvature and Gaussian curvatures as important shape parameters are used as the local descriptions of the 3D surface because these are inherent features of the surface. An important problem of uncertainity management of multiple interpretations of a pair of Images of a surface is proposed to be resolved by developing domain dependent knowledge based expert systems for the particular computer vision system. Finally, a computer vision scheme based upon the above philosophy is presented with some results.