(The) effect of image filtering on optical flow estimation and an efficient filtering technique

Abstract

Gradient-based methods are widely in the computation of optical flow, but subject to two principal problems. The first is inaccurate measurements of the spatial and temporal gradients of image brightness. Secondly, the smoothness constraint underlying those approaches is to create blurring of motion boundaries. This thesis invesgates the methods to cope with these difficulties. For the former problem, performing the error analysis of gradient-based methods, we show that optical flow estimation error is related with the sampling intervals and the higher-order derivatives of image brightness. By examining the effect of those factors on optical flow estimation, we demonstrate the necessity of image filtering prior to the application of optical flow estimation and show that the filtering is applied to the images adaptively to the velocity. From the frequency-domain analysis of optical flow estimation, we present a velocity adaptive spatial filter which its spatial cutoff frequency, for a fixed temporal cutoff frequency, varies adaptively to the velocity. The experimental results for both synthetic and real image sequences demonstrate its promise as a prefilter for optical flow estimation. For the latter problem, based on the observation that the constant velocity constraint underlying a local optimization method creates difficulties at parts where the image brightness changes abruptly such as occluding contours or region boundaries, we try to overcome such difficulties by incoporating the information of discontinuities in image brightness in the local method. By deriving the relation between image structure and variation in optical flow field, we formulate a weighted local optimization method whereby the influence of its neighbor on an observed point is determined according to both the relative distance and the brightness difference between them. The experimental results show that optical flow field is obtained with less blurring in motion boundaries.