Object Tracking via Multi-Task Gaussian-Laplacian Regression

In this paper, we propose a multi-task object tracking algorithm, which is based on an incremental subspace learning method and is denoted as the Multi-Task Gaussian-Laplacian Regression Tracker (MGLRT). Firstly, we model the candidate targets as a mutli-task linear regression by PCA basis vectors. Secondly, considering the complexity of the real noise in a tracking system, we model the noise as an addition of Gaussian and Laplacian noise. In the corresponding optimization problem, we denote by parallel to E parallel to(2,1) and parallel to S parallel to(1,1) the Multi-Task Gaussian and Laplacian noise term addition, since the candidate targets are densely sampled around the current target state, the representations will be linearly dependent. Therefore, in order to improve the performance of our tracker, the representations matrix is expected to be low-rank. Finally, we test our tracking algorithm on challenging videos that include partial occlusion, illumination variation, pose change, background clutter and motion blur. Experimental results show that our proposed approach performs favorably against several state-of-the-art trackers.