基于分类一致性的迁移学习及其在行人检测中的应用

山东大学学报(工学版) ›› 2013, Vol. 43 ›› Issue (4): 26-31.

基于分类一致性的迁移学习及其在行人检测中的应用

于立萍1,2,唐焕玲1,2

1.山东工商学院计算机科学与技术学院, 山东烟台 264005;
2.山东省高校智能信息处理重点实验室(山东工商学院), 山东烟台 264005

收稿日期:2013-05-14 出版日期:2013-08-20 发布日期:2013-05-14
作者简介:于立萍(1971- ),女,山东烟台人,副教授,博士,主要研究方向为机器学习与计算机视觉. E-mail:yulipingguo@gmail.com
基金资助:
国家自然科学基金资助项目(61175053)

Transfer learning model based on classification consensus and its application in pedestrian detection

YU Li-ping1,2, TANG Huan-ling1,2

1. School of Computer Science and Technology, Shandong Institute of Business and Technology, Yantai 264005, China;
2. Key Laboratory of Intelligent Information Processing in Universities of Shandong (Shandong Institute of Business and Technology), Yantai 264005, China

Received:2013-05-14 Online:2013-08-20 Published:2013-05-14

摘要/Abstract

摘要：

利用迁移学习解决在特定场景下尤其是在摄像头静止的监控场景下的行人检测问题,提出基于分类一致性的学习模型。利用Boosting技术从辅助训练集中选择具有正迁移能力的样本,对样本迁移能力给出了基于辅助分类器分类一致性的熵度量方法。对比实验表明,该学习模型能够有效地提高检测率,尤其是在标记样本较少的情况下仍得到了较好的检测效果。

关键词: 分类一致性, 行人检测, 迁移学习, Boosting

Abstract:

Based on the classification consensus, a novel transfer learning model for a scene-specific pedestrian detector especially in video surveillance with stationary cameras was propose. According to boosting technology, the samples showed positive transferability in auxiliary data set were selected and added to the target data set. The entropy-based transferability measurement was derived from the consensus on the predictions of auxiliary classifications. Experimental results showed that the proposed approach could improve the detection rate, especially with the insufficient labeled data.

Key words: transfer learning, Boosting, pedestrian detection, classification consensus

中图分类号:

TP181

于立萍1,2,唐焕玲1,2. 基于分类一致性的迁移学习及其在行人检测中的应用[J]. 山东大学学报(工学版), 2013, 43(4): 26-31.

YU Li-ping1,2, TANG Huan-ling1,2. Transfer learning model based on classification consensus and its application in pedestrian detection[J]. JOURNAL OF SHANDONG UNIVERSITY (ENGINEERING SCIENCE), 2013, 43(4): 26-31.

参考文献

多维度评价

Viewed

Full text

Abstract

Cited

Shared

Discussed