基于双层路由注意力机制的目标跟踪算法  

Object tracking algorithm based on bi-level routing attention mechanism

作　　者：孙逸秋 周冬明[1] 王长城 SUN Yi-qiu;ZHOU Dong-ming;WANG Chang-cheng(School of Information Science and Engineering,Yunnan University,Kunming 650000,China)

机构地区：[1]云南大学信息学院,云南昆明650000

出　　处：《计算机工程与设计》2025年第4期959-965,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62066047,61966037)。

摘　　要：为解决目前基于Transformer的目标跟踪算法采用单一注意力机制,难以有效区分背景和目标的优先级,以及全局注意力操作导致计算负担过重的问题,提出一种基于双层路由注意力的目标跟踪算法。双层路由注意力特征融合模块将单一注意力分为两个层次,上层对图像的区域特征进行相似度筛选,下层对筛选出的区域进行注意力计算。所提算法在LaSOT、GOT-10K、OTB100多个数据集上进行对比实验,实验结果表明,所提算法跟踪表现优良,性能优于现有的多个跟踪器。To address the difficulty of effectively differentiating between background and target priority using the current Transformer-based target tracking algorithm employing a single attention mechanism,as well as the computational burden caused by global attention operation,a target-tracking method based on bi-level routing attention was proposed.A bi-level routing attention feature fusion module that divided a single attention into two layers was introduced.In the upper layer,the region features of the image were initially screened for similarity,while the attention computation was performed on these screened regions in the lower layer.Comparative experiments were conducted on multiple datasets including LaSOT,GOT-10K,and OTB100 to evaluate the proposed algorithm.Experimental results demonstrate its superior performances compared to that of several existing trackers in terms of tracking accuracy.

关 键 词：深度学习 特征融合 目标跟踪 模板更新 注意力机制 多尺度结构 单流框架 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]