An improved Siamese network-based tracking method for UAV video with salt-and-pepper noise

Abstract

In practical application of unmanned aerial vehicle (UAV) monitoring system, the video in visual tracking system inevitably introduces salt-and-pepper (s&p) noise because of the interference from external environment. At this time, it could degrade the performance or lead to tracking failure. However, most existing tracking algorithms focus on the improvement of accuracy and robustness while ignore the video quality enhancement. To this end, this paper proposes a Siamese network-based tracker together with a denoising module to reduce the effect of s&p noise. The proposed tracker firstly adds a simpler version of adaptive median layer into feedforward denoising convolutional neural network to form a video quality enhancer to achieve the noise suppression. Then by means of Siamese network, the tracker extracts the features of preprocessed detection frame and template frame, and the region proposal network (RPN) is utilized to generate the foreground candidate box and its position. Thus, the proposed tracker can maintain the stability and effectiveness of tracking under the influences of the s&p noise. In addition, we construct a noisy DTB70 dataset with genetated noise for experimental validation. Experimental results show that the proposed method can track the target effectively at different noise levels. It is worth mentioning that our proposed method reports promising results on s&p noise even with high levels.