Normalized cross-correlation for spherical images

Recent advances in vision systems have spawned a new generation of image modalities. Most of today's robot vehicles are equipped with onmidirectional sensors which facilitate navigation as well as immersive visualization. When an onmidirectional camera with a single viewpoint is calibrated, the original image can be warped to a spherical image. In this paper, we study the problem of template matching in spherical images. The natural transformation of a pattern on the sphere is a 3D rotation and template matching is the localization of a target in any orientation. Cross-correlation on the sphere is a function of 3D-rotation and it can be computed in a space-invariant way through a 3D inverse DFT of a linear combination of spherical harmonics. However, if we intend to normalize the cross-correlation, the computation of the local image variance is a space variant operation. In this paper, we present a new cross-correlation measure that correlates the image-pattern cross-correlation with the autocorrelation of the template with respect to orientation. Experimental results on artificial as well as real data show accurate localization performance with a variety of targets.