Detection and Estimation of Multiple Disparities by Multi-evidential correlation

ID
TR-93-38
Authors
Esfandiar Bandari and James J. Little
Publishing date
October 1993
Length
21 pages
Abstract

This paper addresses detection and estimation of multiple disparities in motion and stereo using multi-evidential correlation. No a priori knowledge of the presence or the absence of or even the number such disparities is assumed. The procedure utilizes two matching kernels, one based on phase correlation and the other based on a variation of cepstral filtering that provide direct estimates of multiple motion or stereo disparities.

Multi-evidential correlation and the kernels utilized are described and results are presented for motion transparency, occluded boundary and multi-frame analysis of reflected images.

Both kernels were found useful, but phase correlation showed unstable behavior and very broad peaks in the presence of curved surfaces making recognition of multiple disparities difficult. Cepstrum, on the other hand, had very high signal to noise ratio, and provided stable performance thorough all iterations.