Robust visual tracking via weighted incremental subspace learning

A method is proposed for tracking objects in face with varying viewpoints and partial occlusions. A low-dimensional subspace is built to model the appearance of the target. And each image sample is presented as a coefficient vector in the subspace. A collection of image patches are sampled as the candidates of the object image region in the current frame, and their likelihoods of being the object are evaluated resorting to the reconstructions from the corresponding coefficient vectors. The image patch with highest likelihood is selected as the object image region. Finally, the subspace is incrementally updated based on the coefficient vectors, which are assigned with the temporal weights to enhance the adaptability of the tracker. Experimental results show that our proposed method can locate the target accurately even when the target undergoes changes in viewpoints and partial occlusions.