Robust visual tracking for UAVs with dynamic feature weight selection

The low reliability of feature and tracking box detection remains a problem in visual object tracking. Currently, most discriminative correlation filter (DCF) trackers integrate multiple types of features with fixed weights. However, these approaches cannot be adapted to complex scenes when fixed weights are used for features. In addition, the tracking box of traditional DCF trackers lacks scale and aspect ratio adaptability, which can inevitably lead to excessive background noise. To address these problems, we propose a robust tracking method for unmanned aerial vehicles (UAVs) using dynamic feature weight selection. Specifically, we define a feature weight pool that contains multiple weights for different features. In each frame, we select a weight combination with high reliability from the weight pool. This approach is a form of dynamic feature weight selection since the feature weight may be different for each frame. Furthermore, EdgeBoxes is combined with the DSST and can adapt well to the scale and aspect ratio of the tracking box. Extensive experiments based on UAV123@10fps, VisDrone2018-test-dev, and UAVDT show that our tracker is superior to other state-of-the-art trackers. It is noteworthy that the proposed dynamic feature weight selection method can be embedded into any tracking model using multiple features.