Online discriminative dictionary learning for robust object tracking.

The discriminative ability of dictionary learning algorithms plays a crucial role in various computer vision applications, particularly in visual object tracking. In this paper, a novel visual tracking algorithm based on an online discriminative dictionary learning technique is proposed. The proposed method incorporates target and background information into dictionary learning in order to separate the target-of-interest from a cluttered background effectively. The dictionary thus learnt, can ensure that each class-specific sub-dictionary has a good representation of the samples associated with its own class and a poor representation of the other classes. In contrast to other dictionary learning mechanisms, the proposed method also introduces an error term that aims to capture outliers (e.g., noise and occlusion) and minimize its effect on tracking. Furthermore, by optimizing a constrained objective function, the learnt dictionary is rendered robust and discriminative, thereby resulting in an accurate tracking framework that can efficiently separate the target from the background. Finally, an effective and simple observation likelihood function based on the reconstruction errors from both positive and negative templates is designed to achieve better tracking performance. Experimental results on a publicly available benchmark dataset demonstrate that the proposed tracking algorithm performs better than several baseline trackers. (C) 2017 Elsevier B.V. All rights reserved.