Sie befinden Sich nicht im Netzwerk der Universität Paderborn. Der Zugriff auf elektronische Ressourcen ist gegebenenfalls nur via VPN oder Shibboleth (DFN-AAI) möglich. mehr Informationen...
Visual object tracking has been a fundamental topic in recent years and many Siamese structure based trackers have achieved state-of-the-art performance on multiple benchmarks. However, most Siamese trackers default the first frame of the video sequence as the template frame, and the template is not updated or updated linearly in the subsequent tracking process. When the target is rapidly moving, deformed, partially blocked, the tracker is prone to tracking drift. To tackle this issue, in this paper, we propose a novel adaptive template update network (ATUNet), which takes the template of the object at different moments including the initial template, the accumulated template and the predicted template as the input of a frame residual difference module to update the required template of the current frame. Furthermore, we introduce a n-step iterative training to avoid cumbersome and inefficient training process. The ATUNet is compact and can easily be integrated into existing Siamese trackers. We demonstrate the generality of the proposed approach by applying it to three Siamese trackers, SiamFC and SiamRPN and SiamDW. Extensive experiments on VOT2016, VOT2017 datasets demonstrate that our ATUNet effectively predicts the new target template, outperforming the standard linear update.