Visual Tracking Via Sparse Representation With Compressed PCA Basis Vectors

In this paper, we present a novel visual tracking algorithm based on sparse representation. In contrast to the methods using target templates or PCA basis vectors and trivial templates to sparsely represent the object appearance model, which require to solve the l
₁
Regularized Least Square(l
₁
RLS) problem with rather expensive computational cost to obtain the sparse coefficients, we propose an effective representation with a linear combination of compressed PCA basis vectors and trivial templates. First, a sparse measurement matrix is applied to project the candidate image patch into a low dimensional compressed subspace. Then, the object tracking is converted to a sparse approximation problem via compressed features and achieved by solving the l
₁
RLS problem in a Bayesian inference framework. It decreases the computational complexity without loss of tracking accuracy. Furthermore, an inversely indicative strategy is proposed to search the image patch with the largest observing likelihood in original image space according to corresponding result in compressed subspace. Then, different processes are adopted to update the observation model according to the occlusion-ratio of the selected image patch for handling appearance variation. Both qualitative and quantitative evaluations on challenging image sequences demonstrate that the proposed algorithm performs favorably against several state-of-the-art methods in a wide range of tracking scenarios.