A group theory approach to neural network computation of 3D rigid motion

A novel approach is presented to neural network computation of 3D rigid motion. The scheme employs a cost minimization approach based on the assumption of local rigidity. The key to the approach is to designate the cost function in terms of 2D vector fields which represent the infinitesimal generators of the 3D Euclidean group. This approach allows the authors to calculate 3D motion parameters for each image position through a local process. The result of this local process can serve as a base for further perceptual synthesis to delineate larger homogeneous regions of motion. The initial results of a computer simulation of this Lie group-based neural network verifies the approach to 3D motion perception.<>