基于模板特征缓存与自适应注意力的无人机目标跟踪

PDF下载HTML

Title:

UAV object tracking based on template feature buffer and adaptive attention

作者:

陈泷¹, 丁锰^1,2, 石磊³, 黎智辉⁴, 许晓宇⁵, 潘亦伦¹

1. 中国人民公安大学 侦查学院, 北京 100038;
2. 中国人民公安大学 公共安全行为科学实验室, 北京 100038;
3. 中国传媒大学 媒体融合与传播国家重点实验室, 北京 100024;
4. 公安部鉴定中心, 北京 100038;
5. 广东省证据材料司法鉴定工程技术研究中心, 广东 深圳 518033

Author(s):

CHEN Long¹, DING Meng^1,2, SHI Lei³, LI Zhihui⁴, XU Xiaoyu⁵, PAN Yilun¹

1. College of Investigation, People’s Public Security University of China, Beijing 100038, China;
2. Public Security Behavioral Science Lab, People’s Public Security University of China, Beijing 100038, China;
3. State Key Laboratory of Media Convergence and Communication, Communication University of China, Beijing 100024, China;
4. Institute of Forensic Science of China, Beijing 100038, China;
5. Guangdong Provincial Forensic Science of Evidence Materials Engineering Technology Research Center, Shenzhen 518033, China

关键词:

目标跟踪; 无人机; 模板特征缓存; 自适应注意力; 即插即用; 历史信息; 计算机视觉; 深度学习

Keywords:

object tracking; UAV; template feature buffer; adaptive attention; plug-and-play; historical information; computer vision; deep learning

分类号:

TP391.4

DOI:

10.11992/tis.202507009

文献标志码:

2026-2-4

摘要:

现有无人机目标跟踪方法通常仅保留最近帧的模板信息，容易造成重要历史外观信息的丢失。为解决无人机目标跟踪中的目标信息丢失和历史信息有效利用问题，本文提出一种基于模板特征缓存和自适应注意力的无人机目标跟踪方法。1)设计模板特征缓存模块，通过特征缓存区系统性保存多样化的历史目标外观信息，有效解决传统方法中历史信息丢失的问题。2)提出自适应注意力机制，采用通道级注意力动态评估存储特征的重要性，实现对历史模板信息的自适应加权利用。3)采用即插即用架构，可无缝集成到现有主流跟踪器中，增强了算法的实用性和通用性，并设计对称序列评估方法验证历史目标信息的有效保持和利用。实验结果表明，所提方法在5种主流跟踪算法上均取得显著性能提升，在UAV123数据集上AUC平均提升2.34百分点，在UAV20L数据集上提升7.24百分点，在扩展数据集UAV123-L和UAV20L-L上分别提升4.11和9.14百分点，验证了方法的有效性和适用性。

Abstract:

Existing UAV object tracking methods typically retain only the template information from recent frames, leading to the loss of critical historical appearance information. To address the challenges of information loss and effective utilization of historical data in UAV tracking, this paper proposes a novel tracking method based on template feature buffer and adaptive attention mechanisms. 1)we design a template feature buffer module that maintains a comprehensive repository of historical target appearances through a sliding window mechanism, effectively addressing the information loss problem inherent in traditional methods. 2) we introduce an adaptive attention mechanism that employs channel-level attention to dynamically evaluate the relevance of stored features, enabling intelligent weighting of historical template information. 3) we adopt a plug-and-play architecture that integrates seamlessly with existing mainstream trackers, enhancing the algorithm’s practicality and versatility, and design a symmetric sequence evaluation method to validate the effective retention and utilization of historical target information. Experimental results demonstrate significant performance improvements across five mainstream tracking algorithms, with average AUC improvements of 2.34 percentage points on the UAV123 dataset, 7.24 percentage points on the UAV20L dataset, 4.11 percentage points on UAV123-L, and 9.14 percentage points on UAV20L-L, validating the effectiveness and broad applicability of our method.