Ensemble Visual Tracking with Online Multi-view Randomized Trees and Subspace Update *

Visual tracking is still a challenging problem. Tracking drift easily takes place when foreground and background appear similar. In this paper, in order to alleviate the tracking drift problem, we propose a novel visual tracking method by combining multi-view randomized trees and subspace update in an ensemble tracking framework. In our proposed framework, an adaptive multi-view randomized trees is firstly introduced to obtain the accurate confidence map of foreground objects and background. It is the first time that the multi-view randomized trees is introduced into ensemble tracking to improve the accuracy. Secondly, a mean shift tracker detects the object location by seeking the best mode on the confidence map. Moreover, with random forest classifier, the application range of mean shift tracker is also extended from the RGB color space to high-dimensional feature space. To prevent the model from accumulated wrong update, an incremental principal component analysis tracker is implemented as an extra supplement for the ensemble framework to keep the discriminative ability of the tracker. Experimental results have demonstrated that the proposed tracking algorithm consistently provides more powerful ability to decrease the tracking drift than other state-of-the-art approaches.