A recursive least squares solution for recovering robust planar homographies

Abstract

Presented is a recursive least squares (RLS) solution for estimating planar homographies between overlapping images. The use of such a technique stems from its ability in dealing with corrupted and periodic measurements to provide the best solution. Furthermore, its capacity for providing reliable results for time varying parameter estimation also motivates its use in the context of real time cooperative image mosaicing where optimal transformation between mobile platforms is likely to change due to motion and varying ambient conditions and thus a way to tackle this problem real time is what is required. Additionally, and within the same context, a derived "match making" algorithm is introduced based on high curvature points (Harris points) and 3D intensity histograms which are in-turn matched using the L 2 and L ∞ and then compared to classical cross correlation(CC) techniques. Experimental results show that for synthetic data heavily corrupted by noise the RLS does a decent job of finding an improved homography, provided that the initial estimate is good. Results from real image data show similar results where the homography estimate is improved upon by periodic measurements. The match making algorithm proposed fairs well compared to intensity vector techniques, with the L ∞ based method coming out on top. © 2011 Springer-Verlag Berlin Heidelberg.