基于Transformer的尺度自适应高光谱目标跟踪器  

Transformer-Based Scale-Adaptive Hyperspectral Target Tracker

作　　者：黄鲲鹏 魏雨晴 虞洁 窦业天 徐焕宇 赵东 王青 HUANG Kunpeng;WEI Yuqing;YU Jie;DOU Yetian;XU Huanyu;ZHAO Dong;WANG Qing(Wuxi University,School of Electronics and Information Engineering,Wuxi 214000,China;Wuxi University,School of Internet of Things Engineering,Wuxi 214000,China)

机构地区：[1]无锡学院电子信息工程学院,江苏无锡214000 [2]无锡学院物联网工程学院,江苏无锡214000

出　　处：《电光与控制》2024年第12期78-83,共6页Electronics Optics & Control

基　　金：国家自然科学基金(62001443,62105258);江苏省自然科学基金(BK20210064);无锡市创新创业资金“太湖之光”科技攻关计划(基础研究)项目(K20221046);无锡学院人才启动基金(2021r 007);无锡学院优秀本科毕业论文(设计)支持计划(BSZC2023031)。

摘　　要：在高光谱目标跟踪领域,由于目标尺度的不断变化,跟踪目标在连续帧之间呈现多样化的尺度,从而引发跟踪精度下降的问题。为解决跟踪目标尺度变化的问题,提出一种基于Transformer的尺度自适应高光谱目标跟踪器,旨在显著提升高光谱目标跟踪在尺度变化方面的准确性。首先,对高光谱图像进行主成分分析降维操作得到3波段图,然后对搜索图像进行SwinBlock-Crop25的预处理,再将模板图像和预处理后的搜索图像输入到D-swin特征提取模块中,实现多块多层深度特征的提取。随后,对获得的特征进行拼接操作,并应用自注意力机制,以更好地捕捉目标在不同尺度下的关键信息。最终,通过一个多层感知机,将经过自注意力处理的搜索图像特征映射为最终的目标边界框,从而完成目标的跟踪过程。实验结果表明,与现有多个先进的跟踪器进行对比,所提跟踪器具备更高的成功率和准确率,在保持实时性的同时,比现有公认最先进目标跟踪算法MixFormer在应对目标尺度变化挑战上成功率高0.9个百分点。In the field of hyperspectral target tracking,the continuous variation in target scale leads to diverse appearances of tracked objects between consecutive frames,resulting in a decline in tracking accuracy.To address the challenge of target scale variations,a Transfomer-based Scale-adaptive Hyperspectral target Tracker(TSHT)is proposed.The method aims to significantly improve the accuracy of hyperspectral target tracking.Firstly,principal component analysis is applied to hyperspectral images to obtain three-band images.Subsequently,the search images is preprocessed by SwinBlock-Crop25,and then the template image and the preprocessed search image are input into the D-swin feature extraction module for the extraction of multi-block and multi-layer deep features.Following that,the obtained features are concatenated and subjected to a self-attention mechanism to better capture critical information about the target at different scales.Finally,through a multi-layer perceptron,the search image features processed by self-attention are mapped to the final target bounding box,completing the target tracking process.Experimental results demonstrate that TSHT exhibits high success rates and accuracy.In comparison with several state-of-the-art trackers,it outperforms the advanced target tracking algorithm MixFormer by 0.9 percentage points in handling challenges related to target scale variations,while maintaining real-time performance.

关 键 词：高光谱 目标跟踪 TRANSFORMER 图像预处理 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]