Background Modelling on Tensor Field for Foreground Segmentation

摘要: The paper proposes a new method to perform foreground detection by means of background modeling using the tensor concept. Sometimes, statistical modelling directly on image values is not enough achieve good discrimination. Thus may be converted into more information rich form, such as field, yield latent discriminating features. Taking account theoretically well-founded differential geometrical properties Riemannian manifold where tensors lie, we propose approach for field based data Gaussians mixtures domain. We introduced online Kmeans approximation Expectation Maximization algorithm estimate parameters an Affine-Invariant metric. This metric has excellent theoretical but essentially due space curvature computational burden high. novel family metrics, called Log-Euclidean, in order speed up process, while conserving same properties. Contrary affine case, obtain with null curvature. Hence, classical tools usually reserved vectors are efficiently generalized Log-Euclidean framework. Theoretical aspects presented and frameworks compared experimentally. From practical point view, results similar those framework obtained much faster. Theoretic analysis experimental demonstrate promise effectiveness proposed