Object Relocation Visual Tracking based on Siamese network
Jianlong Zhang, Qiao Li, Bin Wang, Chen Chen, Tianhong Wang, Yang Zhou, Ji Li · 2021
Siamese network based trackers treat the visual tracking task as a similarity matching process, which significantly increases the performance of trackers. However, there are still many remaining problems as following. 1) The lightweight backbone of Siamese network limits its feature representation ability, and the tracker is prone to failure under the interference distractors (e.g., similar objects or background) or large viewing angle changes. 2) The single template could not well handle the shape changing of object. 3) Lack of arbiter-corrector mechanism, the tracker cannot workwell once losing the object. To solve these problems, we firstly proposed a novel U-shaped arbiter that could automatically judge whether the tracker fails or not according to a distance histogram. This histogram is built by counting the Euclidean distance between multiple candidate object centers. Secondly, Kuhn-Munkres algorithm is introduced to choose the winner set from a dynamic template set constructed through a backtracking procedure, and relocation of the object is realized by comparing the Structural Similarity between the winner set and the template set. The experiments show that the proposed tracker achieves better performance on visual tracking benchmarks, i.e., VOT2018, OTBIOO, GOT-lOk and LaSOT, against the state-of-the-art methods.