Attentional convolutional neural networks for object tracking

As low-altitude airspace opens up, aeronautical surveillance based Unmanned Aerial Vehicle (UAV) has started to be widely used in the transportation system. Visual object tracking plays an important role in aeronautical surveillance for its accuracy and timeliness. Although traditional trackers have made great progress, they still tend to fail in complex scenes, such as occlusions, illumination variations, background clutter, and etc. In order to make use of appearance features to distinguish the object and surroundings, we propose a novel architecture called attentional convolutional neural networks (ACNN) in conjunction with offline training and online learning for object tracking. ACNN consists of a trunk equipped with attention blocks that highlight the interesting object, and several branches, which are respectively responsible for specific training sequences. In the tracking stage, all branches are removed and a new fully-connected (fc) layer is added to accomplish binary classification. We regard the candidate with the highest probability as current target. Extensive experimental results on public benchmark show that our method performs outstandingly against state-of-the-art methods. In addition, we have also investigated the relationship between the number of network layers and tracking performance for its practical use.