《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (3): 833-843.DOI: 10.11772/j.issn.1001-9081.2021030432

• 人工智能 • 上一篇    

融合层次特征和混合注意力的目标跟踪算法

朱文球1,2(), 邹广1,2, 曾志高1,2   

  1. 1.湖南工业大学 计算机学院,湖南 株洲 412000
    2.湖南省智能信息感知与处理技术重点实验室,湖南 株洲 412000
  • 收稿日期:2021-03-22 修回日期:2021-06-15 接受日期:2021-06-17 发布日期:2022-04-09 出版日期:2022-03-10
  • 通讯作者: 朱文球
  • 作者简介:邹广(1997—),男,湖南岳阳人,硕士研究生,主要研究方向:数字图像处理、目标跟踪
    曾志高(1973—),男,湖南攸县人,教授,博士,主要研究方向:机器学习、数字图像处理、智能计算。
  • 基金资助:
    国家重点研发计划项目(2019QY1604);国家自然科学基金资助项目(U1836217);湖南省教育厅开放平台创新基金资助项目(20K046)

Object tracking algorithm with hierarchical features and hybrid attention

Wenqiu ZHU1,2(), Guang ZOU1,2, Zhigao ZENG1,2   

  1. 1.School of Computer Science,Hunan University of Technology,Zhuzhou Hunan 412000,China
    2.Hunan Province Key Laboratory of Intelligent Information Perception and Processing Technology,Zhuzhou Hunan 412000,China
  • Received:2021-03-22 Revised:2021-06-15 Accepted:2021-06-17 Online:2022-04-09 Published:2022-03-10
  • Contact: Wenqiu ZHU
  • About author:ZOU Guang, born in 1997, M. S. candidate. His research interests include digital image processing, objection tracking.
    ZENG Zhigao, born in 1973, Ph. D., professor. His research interests include machine learning, digital image processing, intelligent computing.
  • Supported by:
    National Key Research & Development Project of China(2019QY1604);National Natural Science Foundation of China(U1836217);Open Platform Innovation Foundation of Hunan Provincial Education Department(20K046)

摘要:

目标跟踪任务中,全卷积孪生网络的目标跟踪(SiamFC)算法在目标遮挡、光照变化等场景时会表现出鲁棒性较差、丢失跟踪目标等问题,为此提出一种结合特征融合和注意力机制的目标跟踪算法。首先,采用ResNet50作为主干网络提取更充分的目标特征;其次,结合注意力机制对特征进行筛选,将筛选后的低层模板特征与高层模板特征分别同对应搜索特征做互相关操作后进行自适应加权融合,提升网络对正负样本的辨别力。在OTB100数据集上测试,所提算法的精度和成功率分别为81.25%和64.06%;在LaSOT数据集上测试,该算法的精度和成功率分别为49.4%和50.1%。实验结果表明,该算法目标跟踪性能优于全卷积孪生网络算法,且在处理复杂场景时有更好的鲁棒性。

关键词: 目标跟踪, 深度卷积神经网络, 层次特征融合, 注意力机制, 孪生网络

Abstract:

In object tracking tasks, Fully-Convolutional Siamese network for object tracking (SiamFC) algorithm has problems such as poor robustness and loss of tracking objects under the scenes of object occlusion and illumination variation. Therefore, an object tracking algorithm combining attention mechanism and feature fusion was proposed. Firstly, ResNet50 (Deep Residual Network) was used as the backbone network to extract more adequate object features. Secondly, attention mechanism was used to filter features. After low-level template features and high-level template features were correlated with the corresponding search features, the adaptive weighted fusion was carried out to improve the discrimination of positive and negative samples. Tested on the OTB100 (Object Tracking Benchmark) dataset, the proposed algorithm had the precision and success rate of 81.25% and 64.06%. Tested on the LaSOT (high-quality benchmark for Large-scale Single Object Tracking) dataset, the proposed algorithm had the precision and success rate of 49.4% and 50.1%. Experimental results show that the object tracking performance of the proposed algorithm is better than that of the fully convolutional Siamese network algorithm, and it has better robustness when dealing with complex scenes.

Key words: object tracking, deep convolutional neural network, hierarchical feature fusion, attention mechanism, Siamese network

中图分类号: