电子学报 ›› 2016, Vol. 44 ›› Issue (4): 868-872.DOI: 10.3969/j.issn.0372-2112.2016.04.016

• 学术论文 • 上一篇    下一篇

一种运动数据检索的相关反馈算法

陈松乐1,2, 孙正兴1, 张岩1, 李骞1   

  1. 1. 南京大学计算机软件新技术国家重点实验室, 江苏南京 210046;
    2. 南京邮电大学宽带无线通信与传感网教育部重点实验室, 江苏南京 210003
  • 收稿日期:2014-02-21 修回日期:2015-06-23 出版日期:2016-04-25
    • 通讯作者:
    • 孙正兴
    • 作者简介:
    • 陈松乐 男,1976年生于江苏淮阴.博士.现为南京邮电大学物联网学院讲师.主要研究方向为多媒体数据分析与处理、计算机视觉. E-mail:chensongle@hotmail.com
    • 基金资助:
    • 国家高科技发展计划 (No.2007AA01Z334); 国家自然科学基金 (No.61272219,No.61100110,No.61321491); 教育部新世纪优秀人才资助计划 (No.NCET-04-0460); 江苏省科技计划 (No.BE2010072,No.BE2011058,No.BY2012190,No.BY2013072-04); 计算机软件新技术国家重点实验室创新基金重点项目 (No.ZZKT2013A12)

A Relevance Feedback Algorithm for Motion Data Retrieval

CHEN Song-le1,2, SUN Zheng-xing1, ZHANG Yan1, LI Qian1   

  1. 1. State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, Jiangsu 210046, China;
    2. Key Laboratory of Ministry of Education of China for Broadband Wireless Communication and Sensor Network Technology, Nanjing University of Posts and Telecommunications, Nanjing, Jiangsu 210003, China
  • Received:2014-02-21 Revised:2015-06-23 Online:2016-04-25 Published:2016-04-25
    • Supported by:
    • National high-tech Development Plan (No.2007AA01Z334); National Natural Science Foundation of China (No.61272219, No.61100110, No.61321491); Program for New Century Excellent Talents in University of Ministry of Education of China (No.NCET-04-0460); Jiangsu Science and Technology Program (No.BE2010072, No.BE2011058, No.BY2012190, No.BY2013072-04); Key Innovative Program of State Key Laboratory for Novel Software Technology (No.ZZKT2013A12)

摘要:

本文提出了一种基于RankBoost的运动数据检索相关反馈算法.该算法具有以下二个方面的特点:首先,以KNN-DTW作为RankBoost集成学习的弱排序器,在适应变长多变量时间序列(Variable-Length Multivariate Time Series,VLMTS)数据的同时,利用RankBoost的集成性与高效性解决相关反馈实时性要求与VLMTS数据计算复杂度高的矛盾;其次,以本文提出的最小化排序经验损失和泛化损失风险作为RankBoost集成学习目标,有效地克服了相关反馈小样本学习环境下的过拟合问题.在CMU动作库上的实验结果验证了该方法的有效性.

关键词: 运动捕获数据, 相关反馈, RankBoost, 排序损失

Abstract:

A relevance feedback algorithm based on RankBoost for content-based motion data retrieval (CBMR) is presented and has two characteristics.First,KNN-DTW is employed as the weak ranker for RankBoost ensemble learning.While adapting to variable-length multivariate time series (VLMTS) data,by taking the advantage of the ensemble and efficiency of RankBoost,it can resolve the conflict between the real-time requirement of relevance feedback and the high computational complexity of VLMTS data.Second,minimizing ranking experience loss and generalization loss risk proposed in this paper are used as the learning objective for RankBoost ensemble learning,which can effectively solve the over-fitting problem caused by small-sample training in relevance feedback.Experimental results on CMU action library verify the effectiveness of the proposed algorithm.

Key words: motion capture data, relevance feedback, RankBoost, rank loss

中图分类号: