Object Tracking Algorithm of Fully-Convolutional Siamese Networks Using the Templates with Suppressed Background Information

Abstract

The current visual object tracking algorithm of Fully-Convolutional Siamese Networks (SiamFC) has good performance of accuracy and frame rate. However, when tracking an object moving in a scene with complex background, the templates applied in SiamFC tend to introduce the excessive background information that may cause interference to the target. In this paper, a method of suppressing the background information in templates is proposed to cope with this problem. On one hand, it reduces the introduction of background information by using adaptive aspect ratio when making templates. On the other hand, it decreases the impact of background information on matching results through Gaussian weighting after the templates inevitably introduce background information. The effectiveness of the proposed method has been experimentally validated without loss of real-time performance. In the comparison experiments, the proposed algorithm has improved the area under curve (AUC) of success plots by 9.07% and 13.31% on OTB2013 dataset and OTB50 dataset, respectively, compared with the original SiamFC under the complex background interference scenarios.