A comparison of subspace methods for accurate position measurement

Abstract

A comparison of the accuracy of visual position measurement in four common subspaces is presented. Principal component analysis (PCA), independent component analysis (ICA), kernel principal component analysis (KPCA) and Fisher's linear discriminant (FLD) are examined for their ability to discriminate positions in a 2D visual subspace. The comparison was done with both constant and varying illumination and random occlusion. It is shown that PCA provides very good overall performance compared with more sophisticated techniques such as ICA, FLD, and KPCA, at a reduced computational complexity.