3D motion estimation using single perspective sparse range data via surface reconstruction neural networks

A regularized surface reconstruction neural network approach to perform robust invariant 2D/3D object recognition and motion estimation is presented. By efficiently embedding the whole 2D/3D image space into a neural network parametric representation, it is possible to elegantly duplicate the human's mental image transform and matching capability in performing the rotating and scaling of objects as suggested by the studies of experimental psychology. The preliminary simulations of applying this technique to invariant 2D target classification and 3D object motion estimation using sparse range data collected from a single perspective view are encouraging.<
>