Low-rank tensor tracking

Visual object tracking is an important step for many computer vision applications. Visual tracking becomes more challenging when the target object observes severe occlusion, lighting variations, background clutter, and deformation difficulties to name a few. In the literature, low-rank matrix decomposition methods have shown to be a potential solution for visual tracking in many complex scenarios. These methods first arrange the particles of the target object in a 2-D data matrix and then perform convex optimization to solve the low-rank objective function. However, these methods show performance degradation in the presence of the aforementioned challenges. Because these methods do not consider the intrinsic structure of the target particles, therefore, the object loses its spatial appearance or consistency. To address these challenges, we propose a new low-rank tensor decomposition model for robust object tracking. Our proposed low-rank tensor tracker considers the multi-dimensional data of target particles. We employ the recently proposed tensor-tensor product-based singular value decomposition and a new tensor nuclear norm to promote the intrinsic structure correlation among the target particles. Experimental evaluations on 20 challenging tracking sequences demonstrate the excellent performance of the proposed tracker as compared with state-of-the-art trackers.