1. Introduction
Using covariance matrices of image features for object tracking has drawn increasing interest lately. [1] uses covariance matrices for characterizing the spatial features, sta-tistical properties and correlations within the similar objects. It enables efficient fusion of different type of features while keeping small dimensionality, and is shown to be robust and versatile for variations in illuminations, views and poses at modest computational cost. The space of non-singular covariance matrices of image features (or, Symmetric Positive Definite (SPD) matrices) can be formulated as connected points on the Riemannian manifold. The Log-Euclidean and affine invariant metrics [3], [2] provide a framework for generating the statistics on the Riemannian manifold. Numerical results of both metrics are similar, however, the first metric has a simpler form of distances and Riemannian means as compared with the second metric that has no closed form solution for Riemannian means.