面向动态复杂场景的红外小目标自适应跟踪方法(特邀)  

Adaptive tracking method for infrared small targets in dynamic and complex scenes(invited)

作　　者：马天磊 刘新浩 彭金柱[1,2] 开志强 王浩 MA Tianlei;LIU Xinhao;PENG Jinzhu;KAI Zhiqiang;WANG Hao(School of Electrical and Information Engineering,Zhengzhou University,Zhengzhou 450001,China;The State Key Laboratory of Intelligent Agricultural Power Equipment,Zhengzhou University,Luoyang 471039,China)

机构地区：[1]郑州大学电气与信息工程学院,河南郑州450001 [2]智能农业动力装备全国重点实验室,河南洛阳471039

出　　处：《红外与激光工程》2025年第3期310-324,共15页Infrared and Laser Engineering

基　　金：国家自然科学基金项目(62373330);河南省高校青年骨干教师资助项目(2023GGJS005);河南省重点研发计划项目(241111110500);河南省科技研发计划联合基金-青年科学家项目(235200810073);河南省高校科技创新人才支持计划项目(25HASTIT029);中原科技创新青年拔尖人才项目。

摘　　要：红外小目标跟踪是红外探测系统研究的热点之一,其在民用和军事上均有广泛的应用。然而,由于红外小目标的目标弱、背景强和动态变化的特点,使得红外小目标跟踪仍然充满挑战。为了解决这些问题,并提高跟踪器的鲁棒性和准确度,文中面向动态复杂场景提出了一种红外小目标自适应跟踪方法。首先,在红外小目标的特征提取阶段,基于孪生网络的概念,提出了一种双通道多尺度特征提取与融合子网络。该子网络旨在最小化特征损失,解决红外小目标图像中的低信杂比(Signal-to-Clutter Ratio,SCR)问题。其次,为了减轻目标变化对跟踪结果的影响,该研究设计了两个模块。一方面,利用网络模型的空间不变性,在模板分支中提出了动态模板特征增强模块(Dynamic Template Feature Enhancement,DTFE)来增强模板特征。另一方面,提出自适应模板更新模块(Adaptive Template Update,ATU),其利用来自历史帧目标的信息来自适应地更新模板特征。两个模块有效缓解了因目标变化引起的跟踪失败。最后,在跟踪支路中提出了多层自注意力模块(Multi-layer Self-attention,MSA),其利用注意力机制来减少背景杂波的干扰。消融实验证明了文中提出的各部分对跟踪结果的贡献。该研究的方法在公开红外小目标序列上的实验结果优于几种先进的方法,其平均成功率和平均精确度分别达到了85.5%和91.5%。与几种先进方法相比,该方法具有更好的成功率和精确度,同时在速度方面满足实时性要求。Objective The tracking of infrared small targets is one of the hot topics in infrared engineering,and it has wide applications in both civilian and military applications.However,due to the weak target,strong background,and dynamic changes of infrared small targets,tracking them remains a challenging task.The main challenge currently faced by infrared small target tracking algorithms is that:1)The signal-to-clutter ratio(SCR)of infrared small target images is low,making the features easily overwhelmed.2)Over time,changes in the scale and attitude of the target can affect tracking precision.3)The presence of background clutter interference in the target environment affects tracking performance.Therefore,it is essential to establish a robust tracking method for infrared small targets.Methods To address these issues and improve the robustness and accuracy of the tracker,this article proposes an adaptive tracking method for infrared small targets in dynamic and complex scenes.Firstly,in the feature extraction stage of infrared small targets,this study proposes a dual channel multi-scale feature extraction and fusion sub network based on the concept of twin networks.This sub network aims to minimize feature loss and solve the problem of low signal-to-noise ratio(SCR)in infrared small target images.Secondly,in order to mitigate the impact of target changes on tracking results,this study designed two modules.On the one hand,this article utilizes the spatial invariance of network models and proposes a dynamic template feature enhancement module(DTFE)in the template branch to enhance template features,thereby obtaining robust initial frame template features and effectively alleviating tracking failures caused by target changes.On the other hand,this article proposes an Adaptive Template Update Module(ATU),which adaptively updates template features using information from historical frame targets,alleviating the impact of target scale changes on tracking results.Finally,this article proposes a multi-layer self attention modu

关 键 词：红外小目标 目标跟踪 杂波抑制 注意力机制 特征增强 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]