Size-aware visual object tracking via dynamic fusion of correlation filter-based part regressors

Abstract Although correlation filter (CF)-based trackers have shown promising results in addressing problematic challenges of visual tracking, common holistic-wise CF-based trackers mostly drift away from the target object when undergoing partial occlusion. On the other hand, part-based models provide a prosperous basis for handling occlusion problem, due to preserving local structure of the target object. Employing local-global appearance models of the object, we propose a robust tracking algorithm based on the weighted cumulative fusion of CF-based part regressors. Indeed, we dynamically learn importance weights of each part via a multilinear ridge regression optimization model aiming at enhancing discrimination power of our tracker. To alleviate tracking drift caused by the object size changes, we further present an accurate method that jointly estimates object scale and aspect ratio by analyzing relative deformation cost of importance pair-wise parts. Also, to reduce the computational complexity, we introduce a feature sharing strategy for all constituent parts. Extensive experiments on OTB-2013, OTB-50, OTB-100, and VOT2016 datasets demonstrate that our tracker not only impressively enhances the performance of target-wise KCF tracker as its baseline but also performs favorably against state-of-the-art trackers in terms of qualitative and quantitative measures while running about 30 fps using Matlab on 3.2 GHz core-i5.