电子学报 ›› 2016, Vol. 44 ›› Issue (5): 1139-1148.DOI: 10.3969/j.issn.0372-2112.2016.05.018

• 学术论文 • 上一篇    下一篇

基于在线随机蕨分类器的实时视觉感知系统

罗大鹏1, 罗琛2, 魏龙生1, 韩家宝1, 王勇1, 马丽1   

  1. 1. 中国地质大学(武汉)机械与电子信息学院, 湖北武汉 430074;
    2. 北京师范大学惠州附属学校, 广东惠州 516002
  • 收稿日期:2014-10-10 修回日期:2015-01-11 出版日期:2016-05-25
    • 通讯作者:
    • 罗大鹏
    • 作者简介:
    • 罗琛 女,1976年7月出生,广东惠州人,毕业于武汉大学,现为北京师范大学惠州附属学校教师,主要研究方向为计算机视觉、图像处理与分析,机器学习等.E-mail:1430319794@qq.com
    • 基金资助:
    • 国家自然科学基金 (No.61302137,No.61271328,No.41202232); 湖北省自然科学基金 (No.2013CFB403); 武汉市晨光计划项目 (No.2014070404010209)

Real Time Visual Perception System Based on Online Fern Classifier

LUO Da-peng1, LUO Chen2, WEI Long-sheng1, HAN Jia-bao1, WANG Yong1, MA Li1   

  1. 1. School of Electronic Information and Mechanics, China University of Geosciences, Wuhan, Hubei 430074, China;
    2. Huizhou School Affiliated to Beijing Normal University, Huizhou, Guangdong 516002, China
  • Received:2014-10-10 Revised:2015-01-11 Online:2016-05-25 Published:2016-05-25
    • Supported by:
    • National Natural Science Foundation of China (No.61302137, No.61271328, No.41202232); Natural Science Foundation of Hubei Province,  China (No.2013CFB403); Wuhan Chenguang Program (No.2014070404010209)

摘要:

本文针对不同成像条件下,目标姿态变化对系统检测性能的影响,提出一种具有自主学习能力的视觉感知系统.该系统能在执行检测任务的同时,通过快速的自主学习提高检测性能,并保持实时目标检测速度.系统包括了目标检测模块及在线学习样本自动获取、标注模块.针对目标检测模块为满足系统自主学习需求,提出随机蕨分类器的在线学习方法,使目标检测模块可持续自我更新,提高检测性能;针对样本自动获取、标注模块则提出最近邻分类器辅助的双层级联标注方法.此外,本文提出自主在线学习框架,整个学习过程不用准备初始训练样本集,通过人工选定一个待检测目标即可进行无需干预的自适应学习,逐渐提高检测性能.实验表明,该方法在多种监控场景中均有较强的自适应能力和较好的目标检测效果.

关键词: 在线学习, 视觉感知, 随机蕨分类器, 目标检测

Abstract:

A novel online learning object detection system is proposed, which can self learning and improve its detection performance wihout human-annotated training data.The system is composed of a object detection module and a sample labeling module.Online fern classifier is used in the object detection module because of its fast online learning speed.Consequentely, our system can learn automatically and detect objects in the real time.Samples, which are used to train the classifier online, are acquired and labeled automatically from a two stages cascade method in the sample labeling module.Instead of training initial classifier from some manual labeled training samples like other online learning detection frameworks, our system can learn automatically after specifying the object to be detected.This can greatly reduce the efforts of labelers.Experimental results on several video datasets are provided to show the adaptive capability and high detection rate of our system.

Key words: online learning, visual perception, fern classifier, object detection

中图分类号: