Single Object Tracking in Satellite Videos with Meta-updater and Knowledge Distillation

The subject of target tracking for satellite videos has received more and more attention. Owing to the development of convolutional neural network, it has been demonstrated that the trackers using deep convolutional features perform well on several object tracking benchmarks. However, it is difficult to achieve excellent results directly in the field of satellite video tracking as there are several challenges such as low resolution, complex backgrounds, limited edge information, many interfering targets, loss of occlusion, etc. Thus, in this paper, we introduce an offline-trained meta-updater into the deep learning tracker to guide the tracker’s update in complex situations. Furthermore, we utilize the idea of knowledge distillation to effectively enhance the feature representation, discrimination and localization abilities in tiny object tracking in satellite video. The experimental results on satellite video datasets show that our tracker outperforms other representative deep convolutional feature-based tracking methods in terms of accuracy and precision.