电子学报 ›› 2019, Vol. 47 ›› Issue (10): 2076-2082.DOI: 10.3969/j.issn.0372-2112.2019.10.008

• 学术论文 • 上一篇    下一篇

基于位置敏感模型的深度跟踪算法研究

查宇飞1, 吴敏2, 库涛3, 陈兵3, 张园强3   

  1. 1. 西北工业大学计算机学院, 陕西西安 710072;
    2. 95894部队, 北京 10200;
    3. 空军工程大学航空工程学院, 陕西西安 710038
  • 收稿日期:2018-06-15 修回日期:2019-04-09 出版日期:2019-10-25 发布日期:2019-10-25
  • 通讯作者: 吴敏
  • 作者简介:查宇飞 男,1979年出生,湖北人.副教授,硕士生导师,主要从事机器学习、计算机视觉、目标跟踪、目标检测及人工智能等方面的研究.
  • 基金资助:
    国家自然科学基金(No.61472442,No.61773397,No.61701524);陕西省科技新星资助(No.2015kjxx-46)

Deep Tracking Algorithm Research Based on Location-Sensitive Model

ZHA Yu-fei1, WU Min2, KU Tao3, CHEN Bin3, ZHANG Yuan-qiang3   

  1. 1. School of Computer Science, Northwestern Polytechnical University, Xi'an, Shaanxi 710072, China;
    2. 95894 Troops, Beijing 10200, China;
    3. Aeronautics Engineering College, Air Force Engineering University, Xi'an, Shaanxi 710038, China
  • Received:2018-06-15 Revised:2019-04-09 Online:2019-10-25 Published:2019-10-25

摘要: 视觉目标跟踪旨在寻找与跟踪目标具有相同语义信息的样本,并在视频中精确定位样本的位置.最近,深度分类模型被用来提取跟踪目标的深度嵌入式特征,然而,由于深度分类模型给予相同类别的样本一样的标签,这样容易导致跟踪模糊,甚至失败.为了解决这个问题,本文将样本的空间位置信息加入深度分类模型中,提出位置敏感损失函数.本文所提出的损失函数不仅继承了分类损失函数的特性,并根据样本的空间位置信息对相同标签的样本进行了排序.也就是说,本文的损失函数可以同时实现类间可分和类内排序.相比于分类损失函数,本文的损失函数更适合目标跟踪任务.本文在OTB100[1]和VOT2016[2]上进行了测试,结果表明本文算法可以实现较好的跟踪性能.

关键词: 目标跟踪, 深度学习, 位置敏感模型, 类内排序

Abstract: Visual target tracking is to find samples that have the same semantic information as the tracking target and pinpoint the position of the sample in the video.Recently,the deep classification model is used to extract the deep embedded features of the tracking target.However,since the deep classification model gives the same class of sample labels,it can easily lead to tracking and even failure.In order to solve this problem,we add the spatial location information of the sample to the deep classification model and propose a location-sensitive loss function.The proposed loss function not only inherits the characteristics of classification loss,but also sorts samples of the same label according to the spatial location information of samples.In other words,the loss function in this paper can also encourage the classification between classes and classes.Compared with the classification loss function,the loss function in this paper is more suitable for the task of target tracking.In this paper,OTB100 and VOT2016 were tested,the results show that this algorithm can achieve better tracking performance.

Key words: target tracking, deep learning, location-sensitive model, intra-class ordering

中图分类号: