Robust object tracking via local constrained and online weighted

Abstract

Accounting for most recent tracking algorithms just only handle one specified challenge, in order to adjust to diverse scenarios in object tracking, we propose a discriminative tracking algorithm based on a collaborative model. In order to account for drastic appearance change, the visual prior have been learned offline by adding the locality regularization term. We transfer the visual prior to represent object and learn a basic discriminative classifier. Next we employ minimal sparse reconstruction error to find the best candidate with the learned classifier. In addition, we derive a parameter update strategy which is based on the candidates' distribution. With this strategy, the candidates' weight can be calculated according to the candidates' distribution online. The tracking is carried out within a Bayesian inference framework with this representation. We use the learned classifier and sparse template to construct the dynamic parameter observation model. Furthermore, the particle filter is used to estimate the tracking result sequentially. Both qualitative and quantitative evaluations on variety of challenging benchmark sequences demonstrate that the proposed tracking algorithm achieves more robust object tracking than the state-of-the-art methods.