摘要
无人机视觉跟踪是无人机应用的核心技术之一。现有无人机跟踪方法对输入搜索区域进行无差别关注学习,导致特征判别力下降,难以应对无人机场景中复杂的背景干扰。本文提出一种聚焦关键信息的目标感知Transformer无人机跟踪器。构建一个集成特征学习和目标搜索的单流跟踪框架,以增强令牌之间的信息交互。提出一种自适应关系建模机制,通过对目标模板和搜索区域令牌进行关系建模和动态分类,提前终止对背景令牌的处理,聚焦关键目标信息。设计了一个特征聚合模块,保留目标的细节特征,增强特征表示的判别力,并引入时序一致性约束以保证特征的稳定性。在UAV123、DTB70和UavDrak135无人机跟踪基准上的实验表明,所提出的算法在无人机跟踪方面达到了较优的性能。
Unmanned aerial vehicle(UAV)visual tracking is a foundational technology in the field of UAV applications.Existing UAV tracking methods focus on the input search area for learning,leading to a decline in feature discrimination and difficulty in dealing with complex background interference in UAV scenarios.This paper proposes a targetaware Transformer UAV tracker that focuses on key information.First,a single-stream tracking framework integrating feature learning and target search is constructed to enhance the information interaction between tokens.Second,an adaptive relationship modeling mechanism is proposed.This mechanism models the relationship between the target template and the search area tokens and dynamically classifies them.As a result,the processing of background tokens is prematurely terminated,and the focus shifts to key target information.A feature aggregation module has been developed to retain the detailed features of the target,enhance the discriminative power of the feature representation,and introduce temporal consistency constraints to ensure the stability of features.Experiments on the UAV123,DTB70,and UavDrak 135 UAV tracking benchmarks demonstrate that the proposed algorithm exhibits superior performance in UAV tracking.
作者
林淑彬
吴贵山
杨文元
LIN Shubin;WU Guishan;YANG Wenyuan(School of Computer Science,Minnan Normal University,Zhangzhou 363000,China;Fujian Province Universities Key Laboratory of Data Science and Intelligence Application,Minnan Normal University,Zhangzhou 363000,China;Fujian Key Laboratory of Granular Computing and Application,Minnan Normal University,Zhangzhou 363000,China)
出处
《智能系统学报》
北大核心
2025年第6期1483-1492,共10页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金青年科学基金项目(12101289)
福建省自然科学基金项目(2022J01891)
福建省教育厅中青年项目(JAT220202)。
关键词
目标跟踪
TRANSFORMER
自适应令牌终止
跟踪框架
特征聚合
无人机
背景抑制
基准
target-tracking
Transformer
adaptive token termination
tracking framework
feature aggregation
unmanned aerial vehicle
background suppression
benchmark
