Bayesian Tracking by Online Co-Training and Sequential Evolutionary Importance Resampling

Object tracking is an indispensable ingredient of many machine vision applications such as intelligent visual surveillance, traffic monitoring, robot and vehicle navigation, human computer interactions, virtual and augmented realities, video compression and indexing, etc. and has drawn considerable attention from computer research communities in recent years. Actually, in the real world scenarios, that is a very challenging task due to the interference of noise, clutters, occlusions, illumination variations and dynamic changes of the object and the background appearance in the complex scene; a quite variety of tracking methods have been proposed to tackle these difficulties in decades (Yilmaz et al., 2006), which can be roughly divided into two categories: the deterministic method and the statistical methods. The deterministic method performs tracking typically by seeking the local extreme of a matching function which measures the similarity between a template and a candidate image; the most widely used similarity measures include the sum of squared differences, the histogram intersection distance, the Kullback-Leibler divergence, the normalized cross correlation coefficient, and the Bhattacharyya coefficient. Some optimization techniques have been proposed to search the local extreme of the matching function such as the meanshift method (Comaniciu et al., 2003) and the optical flow based method (Baker & Matthews, 2004). The drawback of these methods is if the matching function takes into account only the object and not the background, then it might not be able to correctly distinguish the object from the background and tracking might fail. More robust similarity measures are presented recently such as the posterior probability measure (Feng et al., 2008.) and the log likelihood ratio of features (Collins et al., 2005), which takes the background interference into account. Recently, object tracking is treated as a binary classification problem, where the object have to be identified from the background with multiple image cues and better performance over the matching function based approaches such as the template-matching method (Lucas & Kanade, 1981), the view-based method (Black & Jepson, 1998), and the kernel-based method (Comaniciu et al., 2003), etc. was reported in literatures, where a discriminative model for separating the object from the background is trained offline and applied before tracking and termed tracking-by-detection method (Andriluka et al., 2008 ; Avidan 2004; Breitenstein et al., 2009 ; Choudhury et al., 2003; Leibe et al., 2008; Okuma et al., 2004; Wu & Nevatia, 2007). However, to formulate that as an object-background discrimination problem, two important factors need to be treated carefully: what features to choose and how to train the classifiers. Furthermore, since the object and background appearance may change greatly