河北水利电力学院学报

河北水利电力学院学报 ›› 2022, Vol. 32 ›› Issue (1): 1-8.DOI: 10.16046/j.cnki.issn2096-5680.2022.01.001

• 人工智能与机器人专题 •    下一篇

基于注意力机制的Siamese目标跟踪算法研究

张军1,2,刘先禄2,张宇山2   

  1. 1.安徽理工大学,人工智能学院,安徽省淮南市泰丰大街168号 232001;
    2.安徽理工大学,机械工程学院,安徽省淮南市泰丰大街168号 232001
  • 收稿日期:2021-06-24 修回日期:2021-11-18 出版日期:2022-03-31 发布日期:2022-06-16
  • 通讯作者: 刘先禄（1997-），男，四川内江人，在读硕士研究生，主要研究方向为机器人技术。E-mail:2806171875@qq.com
  • 作者简介:张军（1963-），男，福建漳州人，教授，硕导，博士，主要研究方向为机电液一体化、机器人技术；E-mail:zhj63@163.com
  • 基金资助:
    国家创新方法工作专项（2018IM010500）;安徽省科技重大专项计划项目（16030901012）;国家自然科学基金资助项目(51175005)

Research on Siamese Target Tracking Algorithm Based on Attention Mechanism

ZHANG Jun1,2, LIU Xian-lu2, ZHANG Yu-shan2   

  1. 1.College of Artificial Intelligence,Anhui University of Science & Technology,232001,Huainan,Anhui,China;
    2.College of Mechanical Engineering,Anhui University of Science & Technology,232001,Huainan, Anhui,China
  • Received:2021-06-24 Revised:2021-11-18 Online:2022-03-31 Published:2022-06-16

PDF

摘要： 为了进一步提升Siamese神经网络算法在目标跟踪领域的性能，本文对SiamMask的backbone模型基于注意力机制原理进行了重新设计。首先，对Siamese目标追踪网络的backbone网络框架进行局部修改；其次，对改进的算法与原SiamMask算法在Microsoft COCO2017数据集上进行了网络训练与验证；最后，将原SiamMask算法与改进的算法的验证集数据进行比对，以此来评价改进算法的性能。结果表明，在同等算力与数据集的条件下，基于注意力机制的backbone Siamese目标跟踪算法比SiamMask在IOU值上有2个百分点左右的性能提升。

关键词: Siamese, IOU, Microsoft COCO2017, backbone

Abstract: In order to further improve the performance of Siamese for the object tracking,in this paper, the backbone of SiamMask was redesigned on the attention mechanism. Firstly, the frame of the backbone of Siamese was modified. Secondly, the improved algorithm and the original SiamMask algorithm were trained and verified in the Microsoft COCO 2017 dataset. Finally, in the Validation set, the original SiamMask algorithm was compared with the improved algorithm to evaluate the performance of the improved algorithm. The results show that under the condition of the same ability of computing and datasets, the Siamese based on attention mechanism has about 2% improvement of IOU over SiamMask.

Key words: Siamese, IOU, Microsoft COCO2017, backbone

中图分类号: