Fast Inference of Contaminated Data for Real Time Object Tracking

The online object tracking is a challenging problem because any useful approach must handle various nuisances including illumination changes and occlusions. Though a lot of work focus on observation models by employing sophisticated approaches for contaminated data, they commonly assume that the samples for updating observation model are uncorrupted or can be restored in updating. For instance, in particle filter based approaches every particle has to be restored for each frame, which is time-consuming and unstable. In this paper, we propose a novel scheme to decouple the observation model and its update in a particle filtering framework. Our efficient observation model is used to effectively select the most similar candidate from all particles only, by analyzing the principal component analysis (PCA) reconstruction with L1 regularization. In order to handle the contaminated samples while updating observation model, we adopt on an online robust PCA during the update of observation model. Our qualitative and quantitative evaluations on challenging dataset demonstrate that the proposed scheme is competitive to several sophisticated state of the art methods, and it is much faster.