Visual Tracking Using Multi-Layer Cnn Features Based Discriminant Correlation Filters With Foreground Mask

This work deals with visual object tracking. The well known discriminant correlation filter (DCF) based approach is improved by multi-layer CNN features, spatial reliability (through a foreground mask) and conditionally model updating strategy. In the training stage, by calculating a foreground mask using the color histograms, for each chosen CNN layer, a correlation filter is trained under the foreground constraint to construct a weak tracker. In next frame, the tracking position is from the weighting of weak trackers, for which the weights are computed by Hedge method. The response peak and oscillation are both considered to estimate the confidence criteria. The model and weight of each weak tracker are updated only when the tracking is high-confident. We analyze and evaluate our system on OTB-13 dataset, and show that our approach performs superiorly against several state-of-the-art methods.