Search

Abstract Existing convolutional neural network (CNN) based trackers have limited tracking performance in some challenging scenarios such as deformation, background clutter and illumination variation, because the features extracted from a single layer or from a linear combination of multiple layers are insufficient to describe the target appearance. To overcome this problem, we propose a novel tracking algorithm based on interactive multiple model (IMM) framework for better exploring deep features from different layers (IMM_DFT). In this method, we first build measurement models from convolutional layers by applying correlation filters on hierarchical features. Then, to effectively estimate the target state for each layer, we design a hybrid system which consists of the foregoing measurement model and an online learning motion model. Finally, in order to achieve the optimal fusion of the systems for adapting diverse appearance variation of the target and background, an IMM estimator is developed to dynamically adjust the weight of each system using likelihood function and transition probabilities. Extensive experiments on OTB-2013, OTB-2015 and VOT-2016 benchmark databases demonstrate that the proposed algorithm achieves more favorable performance than several state-of-the-art methods. [ABSTRACT FROM AUTHOR]