Experimental Study Efficiency of Robust Models of Lucas-Kanade Optical Flow Algorithms in the Present of Non-Gaussian Noise

Abstract

This paper presents experimental efficiency study of noise tolerance model of spatial optical flow based on LucasKanade (LK) algorithms such as original LK with kernel of Barron, Fleet, and Beauchemin (BFB), confidence based optical flow algorithm for high reliability (CRR), robust motion estimation methods using gradient orientation information (RGOI), and a novel robust and high reliability for LucasKanade optical flow algorithm using median filter and confidence based technique (NRLK) under several NonGaussian Noise. These experiment results are comprehensively tested on several standard sequences (such as AKIYO, COASTGUARD, CONTAINER, and FOREMAN) that have differences speed, foreground and background movement characteristics in a level of 0.5 sub-pixel displacements. Each standard sequence has 6 sets of sequence included an original (no noise), Poisson Noise (PN), Salt&Pepper Noise (SPN) at density (d) = 0.005 and d = 0.025, Speckle Noise (SN) at variance (v) = 0.01 and v = 0.05 respectively which Peak Signal to Noise Ratio (PSNR) is concentrated as the performance indicator.