Motion estimation method for blurred videos and application of deblurring with spatially varying blur kernels

Abstract

Optical flow methods, such as Lucas-Kanade and Horn-Schunck algorithms, are popular in motion estimation. However, they fall short on accuracy when they are applied to blurred videos. Some people utilize hybrid camera system to get a low resolution image to suppress the blurring effect so that more accurate optical flow for blurred high resolution image can be further derived, though in most of the practical environments it may not be feasible to deploy hybrid camera systems from cost perspective. In this paper, we propose a novel approach to estimate motion from a blurred video without the use of hybrid camera system, and to reduce motion blur by calculating its spatially varying blur kernels. Essentially, we first separate moving objects into small regions and use the corners of their boundaries as feature points, and then apply Hierarchical Block Matching Algorithm (HBMA) to track them between frames. Motions of non-corner pixels can therefore be estimated by interpolating the motion of these corner points, which further support the calculation of the spatially varying blur kernels for deblurring purpose. Experimental results demonstrate the effectiveness of proposed method.