Tri-Tracking: Combining Three Independent Views For Robust Visual Tracking

Dong Wang,Gang Yang,Huchuan Lu

DOI: https://doi.org/10.1142/S0219467812500210

2012-01-01

International Journal of Image and Graphics

Abstract:Robust tracking is a challenging problem, due to intrinsic appearance variability of objects caused by in-plane or out-plane rotation and extrinsic factors change such as illumination, occlusion, background clutter and local blur. In this paper, we present a novel tri-tracking framework combining different views (different models using independent features) for robust object tracking. This new tracking framework exploits a hybrid discriminative generative model based on online semi-supervised learning. We only need the first frame for parameters initialization, and then the tracking process is automatic in the remaining frames, with updating the model online to capture the changes of both object appearance and background. There are three main contributions in our tri-tracking approach. First, we propose a tracking framework for combining generative model and discriminative model, together with different cues that complement each other. Second, by introducing a third tracker, we provide a solution to the problem that it is difficult to combine two classification results in co-training framework when they are opposite. Third, we propose a principle way for combing different views, which based on their Discriminative power. We conduct experiments on some challenging videos, the results from which demonstrate that the proposed tri-tracking framework is robust.