电子学报 ›› 2017, Vol. 45 ›› Issue (3): 753-761.DOI: 10.3969/j.issn.0372-2112.2017.03.036

• 学术论文 • 上一篇    下一篇

基于加权重叠率的单目标视觉跟踪评价指标

孙巧, 张胜修, 张正新, 曹立佳, 李小锋   

  1. 火箭军工程大学控制工程系, 陕西西安 710025
  • 收稿日期:2016-05-05 修回日期:2016-09-01 出版日期:2017-03-25
    • 通讯作者:
    • 孙巧
    • 作者简介:
    • 张胜修 男,1963年6月出生于陕西西安.博士,教授,博士生导师,主要从事导航、制导与控制方面的研究.E-mail:zsx1963@aliyun.com.cn
    • 基金资助:
    • 国家自然科学基金 (No.61203189); 陕西省自然科学基金 (No.2015JQ6226)

A Weighted-Overlap Based Metric for Single Visual Object Tracking Evaluation

SUN Qiao, ZHANG Sheng-xiu, ZHANG Zheng-xin, CAO Li-jia, LI Xiao-feng   

  1. Department of Control and Engineering, Rocket Force University of Engineering, Xi'an, Shaanxi 710025, China
  • Received:2016-05-05 Revised:2016-09-01 Online:2017-03-25 Published:2017-03-25
    • Supported by:
    • National Natural Science Foundation of China (No.61203189); Natural Science Foundation of Shaanxi Province,  China (No.2015JQ6226)

摘要:

针对真值标注的歧义性、偏差性问题和具有缩放场景的视觉跟踪应用情况,提出了一种新的视觉跟踪单目标基准评价指标.首先,在重叠率基础上提出了加权重叠率框架;其次,提出了多区域标注方法,通过多区域标注降低标注者歧义性,在具有缩放场景的应用中,通过反演进行多区域标注,使评价更符合应用实际;再次,针对标注的偏差性,提出了多标注融合方法,提高了标注的可信度;最后,将应用于单次跟踪评价的重叠率框架推广到多次跟踪评价,利用加权结果图使评价更具解释性.通过著名评价标准VOT、OTB的真值标注融合实验验证了本文标注规则的准确性;通过在具有缩放场景的视觉跟踪实验和重复实验,与其他跟踪指标的比较验证了本文指标的有效性.

关键词: 视觉跟踪, 评价指标, 重叠率, 注释, 缩放

Abstract:

Aimed at the problems of annotation of ground truth and the application of zooming,a new basic metric for visual tracking evaluation is proposed.Firstly,a weighted-overlap frame is reconstructed based on the traditional overlap.Secondly,we put forward multiple region annotation to decrease the deviation and apply in zooming.Thirdly,a multi-label fusion method is presented to improve the confidence level of the labels.Last but not least,the presented methods are expanded to repeated visual tracking evaluation,where a weighted result chart is utilized to make the evaluation more explanatory.Experimental results show that our annotation rule are more accurate than VOT and OTB,and the proposed metric is more appropriate than other metric.

Key words: visual tracking, evaluation metric, overlap, annotation, zooming

中图分类号: