Article Preview
TopIntroduction
Visual tracking, also known as locating an object over time in a video or a visual sequence, has been widely used in multiple fields, such as intelligent monitoring, human-machine interaction, motion analysis, and vehicle navigation. Most of tracking methods are based on various appearance-learning models, and these models are always built from sample data online or offline to represent the tracked object. In general, appearance models consist of two main categories: generative model (Mei & Ling, 2011; Bao, Wu, Ling, & Ji, 2012; Ross, Lim, & Lin, 2008) and discriminative model (Babenko, Yang, & Belongie, 2011; Grabner, & Bischof, 2006; Avidan, 2007). To be specific, generative tracking methods generate the object appearance model by learning the sample’s distribution and searching for the most similar image regions. Discriminative methods consider tracking as a binary classification problem between the object and the background according to the decision boundary.
In recent years, visual tracking methods based on sparse representation have been applied successfully (Mei & Ling, 2011; Bao, Wu, Ling, & Ji, 2012). According to sparse representation theory, the object can be sparsely and linearly represented by a combination of templates. What is more, these methods can properly solve some classic problems of tracking, such as obscured objects, damaged images, the nonlinear object shape, and low-dimensional manifold structure. However, sparse trackers have two problems: (1) How often do the object templates need update? Indeed, the templates need update constantly due to frequent changing of the object appearances (Matthews, Ishikawa, & Baker, 2004). If the updating is too slow, the templates will not adapt to the morphological changes of the object, and meanwhile, the tracking accuracy will be low. If the updating is too fast, some parts of the background will be easily introduced into the templates, which will thereby result in serious drifting. (2) The sparse representation, which uses only a few observation samples, cannot accurately describe the object appearance. Therefore, sparse representation method is relatively weak in distinguishing the object from the background. Because the object template contains not only the object features but also background information, the complex background environment gives rise to unstable and poor controllable tracking results.
Templates are always updated by replacing (Mei & Ling, 2011; Bao, Wu, Ling, & Ji, 2012) and learning (Li, Shen, & Shi, 2011; Chen, Wang, Wang, Zhang, & Xu, 2011; Wang, Lu, & Yang, 2013; Zhang, & Li, 2011; Mairal, Bach, Ponce, & Sapiro, 2009). When the tracking result is different from the current template, the template will be replaced by the one with higher coefficient or higher weight, which means that it is closer to the appearance of the tracking result. In contrast to these replacing approaches, another way is using multiple subspaces learned to model the target appearance variations during the tracking to deal with abrupt appearance changes. It is always assumed that multiple linear models can approximate the appearance manifold of an object. However, since the appearance manifold is always nonlinear and complex, templates’ learning is easy to over-fit and cannot maintain local relationships of the manifolds.