文献类型：期刊文献

中文题名：基于孪生网络和Transformer的小目标跟踪算法SiamTrans

英文题名：SiamTrans:tiny object tracking algorithm based on Siamese network and Transformer

作者：公海涛[1];陈志华[1];盛斌[2];祝冰艳[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]上海交通大学电子信息与电气工程学院,上海200240

年份：2023

卷号：43

期号：12

起止页码：3733

中文期刊名：计算机应用

外文期刊名：journal of Computer Applications

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2023_2024】；

基金：空间智能控制技术全国重点实验室开放基金课题(HTKJ2022KL502010)。

语种：中文

中文关键词：目标跟踪;小目标;孪生网络;注意力机制;Transformer

外文关键词：object tracking;tiny object;Siamese network;attention mechanism;Transformer

摘要：针对现有小目标跟踪算法的鲁棒性差、精度及成功率低的问题,提出一种基于孪生网络和Transformer的小目标跟踪算法SiamTrans。首先,基于Transformer机制设计一种相似度响应图计算模块。该模块叠加若干层特征编码-解码结构,并利用多头自注意力机制和多头跨注意力机制在不同层次的搜索区域特征图中查询模板特征图信息,从而避免陷入局部最优解,并获得一个高质量的相似度响应图;其次,在预测子网中设计一个基于Transformer机制的预测模块(PM),并利用自注意力机制处理预测分支特征图中的冗余特征信息,以提高不同预测分支的预测精度。在Small90数据集上,相较于TransT(Transformer Tracking)算法,所提算法的跟踪精度和跟踪成功率分别高8.0和9.5个百分点。可见,所提出的算法具有更优异的小目标跟踪性能。
Aiming at the problems of poor robustness,low precision and success rate in the existing tiny object tracking algorithms,a tiny object tracking algorithm,SiamTrans,was proposed on the basis of Siamese network and Transformer.Firstly,a similarity response map calculation module was designed based on the Transformer mechanism.In the module,several layers of feature encoding-decoding structures were superimposed,and multi-head self-attention and multi-head cross-attention mechanisms were used to query template feature map information in feature maps of different levels of search regions,which avoided falling into local optimal solutions and obtained a high-quality similarity response map.Secondly,a Prediction Module(PM)based on Transformer mechanism was designed in the prediction subnetwork,and the self-attention mechanism was used to process redundant feature information in the prediction branch feature maps to improve the prediction precisions of different prediction branches.Experimental results on Small90 dataset show that,compared to the TransT(Transformer Tracking)algorithm,the tracking precision and tracking success rate of the proposed algorithm are 8.0 and 9.5 percentage points higher,respectively.It can be seen that the proposed algorithm has better tracking performance for tiny objects.