Real-time Siamese visual object tracking using attention and anchor-free mechanism
Thang Hoang Dinh, Tuan Do Ngoc, Kien Thai Trung, Long Tran Quoc · Journal of Military Science and Technology · 2022
Trackers based on Siamese have consistently demonstrated superior performance in tracking visual objects. The majority of existing trackers calculate the features of the target template and search image independently and then estimate the target's scale and aspect ratio using either a multi-scale searching scheme or pre-defined anchor boxes. This paper proposed a Siamese attention network for tracking visual objects. An attention fusion mechanism is generated using pixel-level matching of template and search features. The framework proposed is anchor-free, making it both simple and effective. Extensive experiments on visual tracking benchmark VOT2018 and UAV123 demonstrate that our tracker operates at 42 fps and achieves state-of-the-art performance.