Designing of the input filter group of the multichannel optical flow estimate method

Abstract

Spatio-temporal derivative method is one of the most effective methods of determining the velocity of moving objects. This method shows excellent response, but it has a problem that estimated velocity is not accurate where Hessian of the target pattern is small or there is noise on the image. It has been reported that one can obtain more stable velocity vectors by synthesizing the velocity vectors obtained for each channel that is consist of different characteristic filters. However, guidelines for the choice of optimal filter group has not been established yet. Therefore, in this study, we researched to suppress unstable velocity vectors at motion edge by using compact filter group. In particular, we made some filter group which consists of different characteristic (passband) filters, and applying it to various image sequences. Then, we searched for the filter group that can provide stable velocity vectors at motion edge on the image. We compared the results in case where filter coefficient is original and normalized. It was found that stable velocity vector can be obtained by applying normalized coefficient filters.