Clustering ensemble tracking

Abstract

A key problem in visual tracking is how to handle the ambiguity in decision to locate the object effectively using the target appearance model with online update. We address this problem by incorporating sequential clustering and ensemble methods into the tracking system. In this paper, clustering is used for mining the potential historical structure in the parameter space and feature space. Then we fuse multiple weak hypotheses to construct a strong ensemble learner for object tracking. Different from previous methods for updating classifier ensemble in a fixed weak classifier pool frame-to-frame, the proposed ensemble method is taking three weak hypotheses into consideration: spatial object-part view, parameter space view, and feature space view. Specially, spatial object-part view represents the object by a collection of part models that are spatially related (e.g. tree-structure). Meanwhile, analyzing the latent group structure in the parameter space and feature space is essential to take full advantage of the historical data in the tracking process. Therefore, we propose a novel ensemble algorithm that fuses object-part predictor, parameter clustered predictors and feature clustered predictors together. Furthermore, the weights of different views are updated by the relative consistency between weak predictors and final ensemble tracker. The formulation is tested in a tracking-by-detection implementation. Extensive comparing experiments on challenging video sequences demonstrate the robustness and effectiveness of the proposed method.