Densely connected Siamese network visual tracking

Abstract

PurposeVisual object tracking plays a significant role in intelligent robot systems. This study aims to focus on unlocking the tracking performance potential of the deep network and presenting a dynamic template update strategy for the Siamese trackers.Design/methodology/approachThis paper presents a novel and efficient Siamese architecture for visual object tracking which introduces densely connected convolutional layers and a dynamic template update strategy into Siamese tracker.FindingsThe most advanced performance can be achieved by introducing densely connected convolutional neural networks that have not yet been applied to the tracking task into SiamRPN. By using the proposed architecture, the experimental results demonstrate that the performance of the proposed tracker is 5.8% (area under curve), 5.4% expected average overlap (EAO) and 3.5% (EAO) higher than the baseline on the OTB100, VOT2016 and VOT2018 data sets and achieves an excellent EAO score of 0.292 on the VOT2019 data set.Originality/valueThis study explores a deeper backbone network with each convolutional network layer densely connected. In response to tracking errors caused by templates that are not updated, this study proposes a dynamic template update strategy.