Neural network techniques for invariant recognition and motion tracking of 3-D objects

Invariant recognition and motion tracking of 3-D objects under partial object viewing are difficult tasks. In this paper, the authors introduce a new neural network solution that is robust to noise corruption and partial viewing of objects. This method directly utilities the acquired range data and requires no feature extraction. In the proposed approach, the object is first parametrically represented by a continuous distance transformation neural network (CDTNN) which is trained by the surface points of the exemplar object. When later presented with the surface points of an unknown object, this parametric representation allows the mismatch information to backpropagate through the CDTNN to gradually determine the best similarity transformation (translation and rotation) of the unknown object. The mismatch can be directly measured in the reconstructed representation domain between the model and the unknown object