基于注意力机制和不对称卷积的目标跟踪算法

Object tracking method based on attention mechanism and asymmetric convolution

一般孪生网络跟踪算法中目标模板不会更新,模板分支与搜索分支在计算时相互独立,无法进行鲁棒跟踪,使用深度互相关来融合两分支的特征有着容易被干扰物欺骗、激活通道数少、对目标边界的分辨能力较弱,且不能充分受益于大规模的离线训练,为此提出一种基于注意力机制和不对称卷积的目标跟踪算法。设计增强注意力网络增强和传递分支信息。采用不对称卷积来代替深度互相关,使用有效的参数学习如何更好地互相关。所提算法在OTB100、LaSOT、VOT2019上做了对比实验,实验结果表明,所提算法表现较好,性能优于现有的多个先进跟踪器。

As to the general Siamese network tracking algorithm,the target template can not be updated,and the template branch and the search branch are independent in calculation,which can not carry out robust tracking and the depth cross-correlation to fuse the features of two branches,causing the problems of being easily deceived by distractors,having a small number of activation channels,the weak ability to distinguish the target boundary,and not being able to fully benefit from large-scale offline training.To solve the problems,a target tracking algorithm based on attention mechanism and asymmetric convolution was proposed.The augmented attention network was designed to enhance and transfer branch information.Asymmetric convolution was used to replace deep cross-correlation,and effective parameters were used to learn how to better cross-correlation.The proposed algorithm was compared on OTB100,LaSOT,and VOT2019.Experimental results show that the proposed algorithm performs well,and its performance is better than that of several existing advanced trackers.