Online collaboration-based visual tracking for unmanned aerial vehicle with spatial-to-semantic information and multi-recommender voting.

Abstract

ABSTRACT Object tracking plays a crucial role in remote sensing for the unmanned aerial vehicle (UAV). In recent years, deep learning contributes hugely to the visual object tracking, and one typical application is that deep features extracted from convolutional neural networks are widely employed for robust representations of the tracked object, as early layers retain higher spatial accuracy and the latter ones contain more semantic information. However, the potential of deep features as well as their fusion has not been thoroughly achieved. In order to fully utilize multi-level deep features, multiple recommenders based on discriminative correlation filters are constructed in this work and provided with a combination of deep features from different layers. Each recommender tracks the object independently and its reliability is evaluated based on the voting from other recommenders as well as from itself. The result of the recommender evaluated as the best will be learned by others adaptively. Extensive experiments on 100 challenging UAV image sequences have demonstrated that the proposed method outperforms recently developed 25 state-of-the-art trackers in terms of robustness and accuracy.