A neural network approach to 3D object identification and pose estimation

A multistage concurrently processing artificial neural network is proposed to identify 3D unoccluded objects from arbitrary viewing angles and to estimate their poses. 3D moment invariants are used to generate feature vectors from 2-1/2D range images. Objects are recognized via moment invariants which are invariant to translation, scaling, and rotation. The proposed network is divided into two stages, the feature extraction stage and the feature detection stage, to generate moment invariants and detect the input features, respectively. Experimental results show that objects coded by 3D moment invariant features can always be satisfactorily classified and estimated by the proposed neural network