Robust Visual Tracking by Hierarchical Convolutional Features and Historical Context

In this paper, we present a visual tracking method to address the problem of model drift, which usually occurs because of drastic change on target appearance, such as motion blur, illumination, out-of-view and rotation. It has been proved that the hierarchical convolutional features of deep neural networks learned by huge classification datasets are generic for other task and can aid the tracker’s power of discrimination. Ensemble-based trackers have been studied also to offer historical context for drift correction. We combine these two advantages into our proposed tracker, in which correlation filters are learned by hierarchical convolutional features and preserved as snapshots in an ensemble in certain occasion. Such an ensemble is capable of encoding the target appearance as well as provide historical context to prevent drift. Such context is considered to be complementary to correlation filters and convolutional features. The experimental results demonstrate the competitive performance against state-of-the-art trackers.