Motion estimation using a neural network

A neural system approach for the implementation of a proposed motion-estimation algorithm is presented. The two-dimensional translational displacement vector is assumed to be uniform among all observables in the area of interest. This assumption simplifies the analysis, and reduces the motion estimation problem to one of image registration. The displacement vector is calculated by relating the maximization of a proposed similarity measure in the image-registration algorithm to the minimization of the network energy function in the neural implementation. The similarity criterion incorporated in the image-registration algorithm uses a coincident bit counting (CBC) method to obtain the number of matching bits between the frames of interest. The CBC method performs favorably compared with traditional techniques, and also renders simpler implementation in conventional computing machines