电子学报 ›› 2021, Vol. 49 ›› Issue (1): 183-191.DOI: 10.12263/DZXB.20180045

• 学术论文 • 上一篇    下一篇

基于少量样本学习的多目标检测跟踪方法

罗大鹏1, 杜国庆1, 曾志鹏1, 魏龙生2, 高常鑫3, 陈应4, 肖菲4, 罗琛5   

  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. 中国地质大学(武汉) 自动化学院, 湖北武汉 430074;
    3. 华中科技大学自动化学院, 湖北武汉 430074;
    4. 中建三局智能技术有限公司, 湖北武汉 430070;
    5. 北京师范大学惠州附属学校, 广东惠州 516002
  • 收稿日期:2018-11-26 修回日期:2019-01-09 出版日期:2021-01-25 发布日期:2021-01-25
  • 通讯作者: 罗大鹏
  • 作者简介:曾志鹏 男,1993年生,湖北仙桃人,中国地质大学(武汉)机械与电子信息学院硕士研究生.主要研究方向为计算机视觉、机器学习.E-mail:zengzhipeng@cug.edu.cn;杜国庆 男,1993年生,湖北武汉人,中国地质大学(武汉)机械与电子信息学院硕士研究生.主要研究方向为发育神经网络、自主学习、目标跟踪与检测.E-mail:guoqingdu@foxmail.com
  • 基金资助:
    国家自然科学基金(No.61302137,No.61603357,No.61603354);中央高校基本科研业务费专项(No.CUGL170210)

Multi-Object Detection and Tracking Based on Few-Shot Learning

LUO Da-peng1, DU Guo-qing1, ZENG Zhi-peng1, WEI Long-sheng2, GAO Chang-xin3, CHENG Ying4, XIAO Fei4, LUO Chen5   

  1. 1. School of Electronic Information and Mechanics, China University of Geosciences, Wuhan, Hubei 430074, China;
    2. School of Automation, China University of Geosciences, Wuhan, Hubei 430074, China;
    3. School of Automation, Huazhong University of Science and Technology, Wuhan, Hubei 430074, China;
    4. Intelligent Technology Co., Ltd. of Chinese Construction Third Engineering Bureau, Wuhan, Hubei 430070, China;
    5. Huizhou School Affiliated to Beijing Normal University, Huizhou, Guangdong 516002, China
  • Received:2018-11-26 Revised:2019-01-09 Online:2021-01-25 Published:2021-01-25
  • Supported by:
     

摘要: 视频目标检测跟踪算法一直是计算机视觉领域的研究热点,目前大部分方法均需人工采集样本训练检测模型,搭建目标检测跟踪系统.当目标成像条件发生变化时,需重新采集样本,训练模型,调试整个检测跟踪系统,耗费大量人力、物力.本文提出一种基于少量样本学习的多目标检测跟踪算法,只需在监控视频第一帧指定待检测目标,即可自主生成混合分类模型,进行目标检测.采用在线渐进学习算法学习目标姿态变化,更新该模型.结合基于颜色的目标跟踪算法,自动构建高精度目标检测跟踪系统.整个过程无需手工采集、标注训练样本.因此,易于扩展到其它监控场景,通过自主学习形成该场景专用的检测跟踪系统,实现不同监控环境下,不同成像条件下都有较好的检测跟踪效果.实验表明,本方法能自主学习多种监控场景下的目标姿态,无需手工标注训练样本,在基于在线学习的算法中有最佳的检测精度,此外也取得了和离线目标检测跟踪系统相似的性能.

 

关键词: 少量样本学习, 多目标检测, 多目标跟踪, 在线学习

Abstract: Video object detection and tracking algorithms have become the research focus in the field of computer vision.Traditional methods need to manually collect samples to train detection models,and build object detection and tracking systems.When the imaging conditions change,it is necessary to re-collect samples to train the detection model and re-adjust the entire detection and tracking system,which requires tedious human efforts.In this paper,a multi-object detection and tracking algorithm is proposed based on few-shot learning.With this approach,a hybrid classifier that models one object class can be generated by simply marking several bounding boxes around the object in the first video frame.An online gradual learning algorithm is proposed to learn the object pose changes and update the model.Combined with the color-based object tracking algorithm,our method automatically builds high-precision object detection and tracking systems without manual collection and labeling training samples.This approach can be conveniently replicated many times in different surveillance scenes and produce scene-specific detectors under various camera viewpoints.Experimental results on several video datasets show our approach achieves comparable performance to robust supervised methods,and outperforms the state-of-the-art online learning methods in varying imaging conditions.

Key words: few-shot learning, multi-object detection, multi-object tracking, online learning

中图分类号: