An Anti-occlusion Correlation Filtering Tracking Algorithm for UAV

In the scenario of Unmanned Aerial Vehicle (UAV), the angle and the height at which the UAV observes the object cause partial occlusion, deformation, and size changing of object image. Based on the Discriminative Correlation Filter (DCF) algorithm, this paper proposes a new tracking algorithm DCF-GA (Discriminative Correlation Filter with Generation of Adversarial example) to achieve anti-occlusion in the scenario of UAV. Firstly, we design a mask selection strategy to generate the adversarial example with occlusion, which can enhance the antiocclusion performance of our algorithm. The response losses of DCF reveal the impacts of adversaries so that they are used to select an appropriate mask. And then, we provide an optimization scheme of object feature selection based on the singular values extracted from histogram of oriented gradient (HOG) feature and convolutional neural network (CNN) feature respectively. Moreover, to overcome the scale changes of the object image, a multidimensional templates set is proposed and the best one is determined by the maximum of their DCF responses. Finally, we add the background patches around the region of interest (ROI) into the sample set to suppress the background clutter. The tracking algorithm we proposed in this paper is compared with some other algorithms in both the UAV video sequence and the OTB dataset. The experimental results show that our DCF-GA algorithm is effective when the object is partially occluded and when the size of object image changes.