Robust object tracking based on selected discriminative convolutional features

Abstract

Video object tracking has been a challenging task in computer vision based applications. Most of state-of-the-art tracking methods rely on convolutional neural network to extract features, and then employ observation model to locate target. Recent studies indicate that convolutional feature maps are noisy and much of the activations are not related to tracking task. Moreover, it will increase computation complexity and undermine feature discriminant ability if raw convolutional features are used to train and update tracking model directly. This paper investigates the impact of refining convolutional features in discriminative correlation filter tracking framework. We employ affinity propagation to search for feature centers from abundant convolutional feature maps, and then take advantage of the fearure centers from different convolutional layers to jointly determine target location. Experimental results on public available dataset show competitive accuracy and robustness against some state-of-the-art tracking methods.