PROJECT TITLE :

An Efficient Tracking System by Orthogonalized Templates

ABSTRACT:

Sparse illustration (SR)-based mostly tracking systems became common within the past recent years for its effectiveness. But, the underlying assumption of those tracking systems is that the target look will be linearly represented by a sparse approximation over a set of templates and a residual term, that typically wants to resolve ℓ1 norm minimization for many times and brings a heavy computational cost. This paper introduces an economical tracking system by discovering orthogonalized templates. By orthogonalizing templates from previous frames and removing their correlation, we show that the sparsity of template weights is not necessary in target look modeling and therefore a least squares regularization will be used. We have a tendency to also decompose the residual term into 2 components in observation model to take occlusion cases into thought. We have a tendency to demonstrate that, in comparison with the SR-primarily based tracking systems that use ℓone learning, our tracking system is much more computationally economical whereas getting a good higher performance. Experiments on a variety of difficult video sequences demonstrate both the effectiveness and potency of the system.