作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (6): 103-107. doi: 10.19678/j.issn.1000-3428.0054400

• 人工智能与模式识别 • 上一篇    下一篇

基于双孪生网络的自适应选择跟踪系统

张腾飞a,b, 周书仁a,b, 彭建a,b   

  1. 长沙理工大学 a. 综合交通运输大数据智能处理湖南省重点实验室;b. 计算机与通信工程学院, 长沙 410114
  • 收稿日期:2019-03-27 修回日期:2019-04-28 发布日期:2019-05-16
  • 作者简介:张腾飞(1993-),男,硕士,主研方向为机器学习、模式识别;周书仁,副教授、博士;彭建,副教授、硕士。
  • 基金资助:
    国家自然科学基金青年基金项目“基于深度神经网络的实体关系抽取关键技术研究”(61602059)。

Adaptive Selective Tracking System Based on Twofold Siamese Network

ZHANG Tengfeia,b, ZHOU Shurena,b, PENG Jiana,b   

  1. a. Hunan Provincial Key Laboratory of Intelligent Processing of Big Data on Transportation;b. Computer and Communication Engineering Institute, Changsha University of Science and Technology, Changsha 410114, China
  • Received:2019-03-27 Revised:2019-04-28 Published:2019-05-16

摘要: 孪生网络在解决目标跟踪问题时具有较大的速度和精度优势,在跟踪领域得到广泛应用。双孪生网络由独立的语义和外观2个分支组成,每个分支都是一个相似学习的孪生网络,解决了原孪生网络精度不足的问题,但其每个分支独立训练,导致系统速度较低。为此,在双孪生网络的基础上提出一种自适应选择跟踪系统ASTS。在测试过程中,简单帧时自动停止网络向前传播,快速判断目标所在位置,从而提高系统的跟踪速度。复杂帧时2个分支相互协调以准确跟踪目标。在OTB2013/50/100和VOT2017数据集上的实验结果表明,相对于固定的双孪生网络目标跟踪方法,ASTS系统具有更快的速度和更高的跟踪准确率。

关键词: 卷积神经网络, 目标跟踪, 孪生网络, 语义信息, 自适应选择

Abstract: Siamese network is widely used in the field of target tracking because of its significant advantage in high speed and accuracy.The twin network is composed of two independent branches:semantic branch and appearance branch.Each branch is a twin network with similar learning,which solves the problem of insufficient accuracy of the original twin network.However,each branch is trained independently,which results in the decrease of system speed.To address this problem,this paper proposes ASTS,an adaptive selective system based on twofold siamese network.In the testing process,the network automatically stops propagating forward at the simple frame and rapidly judge the position of the target,so as to improve the tracking speed of the system.In the case of complex frames,the two branches coordinate with each other to track the target accurately.Experimental results on the OTB2013/50/100 and VOT2017 datasets show that compared with the fixed twofold siamese network object tracking method,the ASTS system has faster speed and higher tracking accuracy.

Key words: Convolutional Neural Network(CNN), object tracking, siamese network, semantic information, adaptive selection

中图分类号: