A Robust and Effective Tracking Method in Remote Sensing Video Sequences
Fukun Bi, Mingyang Lei, Sun Jiayi · 2019
With the popularization of high resolution imaging technology and the progress of artificial intelligence, remote sensing target tracking in the aerial video plays a very important role in public security, such as antiterrorism efforts and military reconnaissance. As aerial video has rapid changes in orientations, low resolution, and multiple similar disruptors, and the main tracking methods generally have relatively low tracking performance in this research field, we develop a robust tracking method for remote sensing videos based on a saliency enhanced multi-domain convolutional neural network (SEMD). The process can be divided into two main stages: (1) in the offline pretraining stage, we combine the Least Squares Generative Adversarial Networks (LSGANs) with a rotation strategy to augment typical easily confused negative samples, which can improve the capacity to distinguish between target and the background. (2) in the online tracking process, a saliency module is embedded between convolutional layers and we optimize the arrangement of its functional sub-modules to boost the saliency of the feature map, which improve the network representation power for rapid dynamic changes in the target. Comprehensive evaluations of homemade datasets demonstrate that the proposed method can achieve high efficiency and accuracy results compared to state-of-the-art methods.