Siamese Region Proposal Network with Multi-level Attention Mechanism and Template Update for Object Tracking
Yang Cao · 2023
Recently, Siamese-based tackers have achieved promising performance and drawn great attention in visual object tracking. However, the tracking algorithms only utilize the deep semantic features of the Siamese network, and the target template is kept fixed without the online update, which is prone to drift in the presence of similar background distractors and large appearance variations. In this paper, we propose a Siamese region proposal network with a multi-level attention mechanism and an online template update strategy (MATU-RPN) to generate more robust and representative features as well as promote adaptation and generalization capabilities. Specifically, a channel and spatial attention module and a feature fusion module are employed to extract discriminative feature representations and combine the beneficial information from multiple layers. Moreover, we construct a template pool and adopt an online update scheme to exploit the underlying temporal relationship of the target object in different frames, which enhances the ability to cope with changeable and complex tracking scenarios. In experiments on OTB100, VOT2018, VOT2019, and LaSOT, MATU-RPN achieves state-of-the-art performance and runs in real-time.