基于Transformer的轻量级单目标跟踪算法  

Lightweight single-object tracking algorithm based on Transformer

作　　者：黄丹丹 张钰晨 陈广秋[1] 刘智[2] HUANG Dandan;ZHANG Yuchen;CHEN Guangqiu;LIU Zhi(School of Electronics and Information Engineering,Changchun University of Science and Technology,Changchun 130022,China;National and Local Joint Engineering Research Center of Space Photoelectric Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学电子信息工程学院,吉林长春130022 [2]长春理工大学空间光电技术国家地方联合工程研究中心,吉林长春130022

出　　处：《华中科技大学学报(自然科学版)》2025年第3期41-47,158,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技厅重点研发项目(20230201071GX)。

摘　　要：针对复杂背景下目标跟踪算法成功率低等问题,提出了基于Transformer的轻量级单目标跟踪算法.算法采用在线时序自适应卷积提取目标局部特征,同时引入轻量级全局上下文模块提取全局特征,共同构建高效的目标模型.为了应对相似图细化中目标信息丢失的问题,构建了轻量特征增强模块,在不增加整体网络参数量的同时增强模型对目标的表达能力,从而提升模型的精确度.最后,算法增加了边界框细化模块,能够更好地保留目标的边界和尺度信息,提高跟踪的准确性.实验结果表明:在UAV123数据集上,相较于基准算法,本文方法的跟踪精确度和成功率分别提高了2.5%和6.9%;在LaSOT数据集上,跟踪精确度和成功率分别提高了8.8%和5.9%;在OTB100数据集上,跟踪精确度和成功率均提高了2.4%.Lightweight single-object tracking algorithm based on Transformer was proposed to address the problems of low accuracy performance of target tracking algorithms in complex contexts.The algorithm adopted online time-series adaptive convolution to extract local features of the target,and at the same time introduced a lightweight global context module to extract global features,which jointly constructed an efficient target model.In order to cope with the problem of loss of target information in the similarity graph refinement,a lightweight feature augmentation module was constructed,which enhanced the model's ability to express the target without increasing the number of parameters of the overall network,so as to increase the accuracy of the model.Finally,the algorithm added a bounding box refinement module,which can better retain the boundary and scale information of the target and improve the tracking accuracy.The experimental results show that compared with the benchmark algorithm,the tracking accuracy and success rate of this paper's method are improved by 2.5%and 6.9%on the UAV123 dataset,8.8%and 5.9%on the LaSOT dataset,and 2.4%and 2.4%on the OTB100 dataset,respectively.

关 键 词：目标跟踪 TRANSFORMER 轻量级 特征增强 边界框细化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]