Motion blur removal for humanoid robots

Abstract

Removing motion blur caused by camera shake is a tough problem which received much attention in past decades. While, blur removal for the images captured by the camera on humanoid robot is more difficult because of the heavy shaking and unpredictable movement at each pace. To account for this challenging blur problem, we propose a hybrid image deblurring algorithm in this paper. Specifically, the images blurred by robot movement are classified as less blurred and severely blurred by using Just Noticeable Blur Metric (JNBM) as a quantitative criterion. For less blurred images, we propose a maximum a posteriori (MAP) framework by taking advantage of the previous sharp image as reference. For severely blurred images, since most details are lost and hard to recover by deconvolution, we refer to the previous neighboring less blurred images, and directly warp the better deblurred one by SIFT matching as the deblurred result. Experimental results demonstrate the proposed algorithm is superior over the existing methods both qualitatively and quantitatively.