A surface-based approach to 3-D object recognition using a mean field annealing neural network

Abstract

Three-dimensional (3-D) object recognition identifies objects in an input image using a modelbase. We present a 3-D object recognition system, in which a symbolic description of the object is generated from the input range data, in terms of the visible surface patches. The segmented surface representation of an object is regarded as a graph, in which nodes contain the information about the individual surface patches whereas links represent the relationships between them such as connectivity. Object recognition is achieved by matching the object graph of a scene with the model graph of a modelbase. In the proposed approach, a mean field annealing (MFA) neural network (NN) is employed as a constraint satisfaction network, in which an energy function robust to occlusion is proposed for effective matching. The energy level is at its minimum when the optimal match is achieved. Simulation results with synthetic and real range images having various occlusion rates are presented to show the effectiveness of the proposed algorithm.