Robust visual tracking based on Gabor feature and sparse representation

This paper proposes a novel approach to deal with the problem of visual tracking in some challenging situations. In our approach, Gabor features of image are used for expressing the templates and candidate targets in order to enhance the robustness of the variations due to illumination and appearance changes. Then we cast tracking as a sparse approximation problem in a particle filter framework. Gabor features derived from the Gabor wavelets representation of image are robust to changes in illumination and expression of the target object. At the same time, the sparse representation is able to deal with the problem of noise, varying viewpoints, background clutter, and illumination changes. The sparse representation is achieved by solving the ℓ1-regularized least square problem. The candidate target with the smallest residual error is considered as the target we want. Most of existing algorithms are unable to track objects for a long time because of the even-changing target and background. In order to overcome the drawback, the template set is renewed by using the incremental learning algorithm which is based on principal components analysis(PCA). We use our approach and other popular methods to track 4 challenging video sequences in which the target objects and the backgrounds change intensively and the targets are partially occluded sometimes. The results show that our method has more excellent performances compared with other methods.