Tensor-SIFT Based Earth Mover’s Distance for Contour Tracking

Abstract

Contour tracking in complex environments is a difficult problem due to the cluttered backgrounds, illumination changes, occlusion and camera viewpoint variations etc. This paper presents a region functional based on the Earth Mover's Distance (EMD), computation of which is mathematically modeled as the transportation problem (TP), for robust contour tracking in the challenging conditions. Formulation of EMD-based functional can be described as variational EMD (VEMD) since the contour curve function is involved for optimization. Minimizing the EMD-based functional is nontrivial and we develop a two-phase method for its optimization. In the first phase, letting the candidate contour be fixed, we seek the best solution to the TP by the Simplex algorithm. Then through the shape derivative theory, we make a perturbation analysis of the contour around the best solution to the TP. As a result we obtain a partial differential equation (PDE) that is solved by the level-set algorithm. The two-phase procedure iterates until the appropriate stopping criterions are satisfied. Alongside the EMD-based functional formulation, we introduce a dimensionality reduction method by tensor decomposition, achieving a low-dimensional Tensor-SIFT features for object representation. Applicable to both the color and gray-level images, Tensor-SIFT features are distinctive, insensitive to illumination and viewpoint changes. Finally, we develop an integrated algorithm that combines various techniques, e.g. the Simplex algorithm, narrow-band level set and fast marching algorithms. Particularly, we provide a method for the level-set initialization between two successive frames and the criterions for stopping the iterative functional optimization. Experiments in challenging image sequences show that the proposed work has promising performance.