SIA OpenIR  > 光电信息技术研究室
目标跟踪中的在线学习方法研究
其他题名Research on Online Learning Methods for Object Tracking
李义翠1,2
导师亓琳
分类号TP391.41
关键词目标跟踪 在线学习 半监督学习 自训练 自步学习
索取号TP391.41/L36/2017
页数63页
学位专业控制工程
学位名称硕士
2017-05-24
学位授予单位中国科学院沈阳自动化研究所
学位授予地点沈阳
作者部门光电信息技术研究室
摘要目标跟踪技术涉及到图像处理、人工智能、模式识别、自动控制等领域,在智能交通、人机交互、智能视频监控、军事装备等领域有着广泛的应用。本文在综述目标跟踪技术的研究背景与现状后,重点研究基于在线学习的目标跟踪方法,并且提出了两种鲁棒的基于在线学习的目标跟踪方法。本文主要工作如下: 1.为解决基于自训练框架的跟踪算法经常发生目标跟踪漂移的问题,提出一种结合PN约束的在线半监督boosting目标跟踪算法(简称PN-SemiT)。自训练框架中用分类器的预测结果来更新分类器,这样一旦有错误引入就会使分类器错误不断累积。PN-SemiT算法引入PN约束对未标记示例的标记进行约束和控制,通过及时纠正被分类器错误分类样本的标签来降低分类器的错误率。此外,该算法还将目标先验知识与在线分类器相结合,在有效改善目标跟踪漂移的同时提高了算法对目标外观变化的适应性。对标准测试集的实验结果表明:PN-SemiT算法具有良好的跟踪性能。 2.为解决长时间目标跟踪中被观测目标易发生遮挡,甚至短时间消失再出现的问题,提出一种基于TLD框架的长时间目标跟踪算法。该算法将跟踪器与分类器相结合,采用中值光流跟踪器实现帧与帧之间的目标跟踪,并且用SVM构建的分类器对图像进行目标检测,最终综合跟踪和检测的结果确定目标位置。此外,跟踪过程中该算法还采用自步学习方法,以回顾性方式从“可靠的”视频帧中提取训练样本,构建包含有丰富负样本的训练集来更新目标模型与分类器。在标准测试集上的实验结果表明:该算法在抗遮挡与目标丢失重捕方面具有较好的性能,同时对尺度和光照变化也有很好的适应性,对复杂环境下长时间运动目标的跟踪具有一定的鲁棒性。
其他摘要Object tracking technology involves image processing, artificial intelligence, pattern recognition, automatic control and other fields. It is also widely used in traffic safety, human-computer interaction, video surveillance, public safety management and military equipment. After reviewing the research background and current situation of object tracking technology, this paper focuses on the object tracking method based on online learning, and proposes two kinds of robust object tracking methods based on online learning. The main work of this paper is as follows: 1. In order to solve the problem of frequent target drift based on self-training framework, an on-line semi-supervised boosting object tracking algorithm (PN-SemiT) is proposed. The so-called self-training is to use the predictor of classifier to update the classifier itself, so that once the introduction of errors will make the classifier errors continue to accumulate. The PN-SemiT algorithm introduces the PN constraint to constrain and control the mark of the unlabeled example, and the error rate of the classifier is reduced by correcting the label of samples. In addition, the algorithm combines a priori knowledge of object and the online classifier to improve the adaptability of the algorithm to the target appearance while effectively improving the target drift problem. Experiments on the standard test set show that the PN-SemiT algorithm has good tracking performance. 2. In order to solve the problem that the observed object in the long-term object tracking is obscured and even disappears for a short time, a long-term object tracking algorithm based on TLD framework is proposed. The algorithm combines the tracker and the classifier, uses the median optical stream tracker to achieve the object tracking between two consecutive frames, and uses the classifier constructed by SVM to carry on the target detection to the image, finally completes the comprehensive tracking result and the detection result to determine the object position. In addition, the algorithm also uses the self-step learning method to retrospectively test the "reliability" of all the frames before this, pick out the "reliable" video frame and extract the training samples from it, then build a training set containing rich negative samples to update target model and classifier. Experiments on the standard test set show that the proposed algorithm has good performance in the field of anti-occlusion and target loss, and has good adaptability to scale and illumination changes. In terms of overall performance the proposed algorithm outperforms other state-of-the-art algorithms, and can achieve robust tracking of long-term moving targets in complex environment.
语种中文
产权排序1
文献类型学位论文
条目标识符http://ir.sia.cn/handle/173321/20544
专题光电信息技术研究室
作者单位1.中国科学院沈阳自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
李义翠. 目标跟踪中的在线学习方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
目标跟踪中的在线学习方法研究.pdf(3145KB)学位论文 开放获取CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[李义翠]的文章
百度学术
百度学术中相似的文章
[李义翠]的文章
必应学术
必应学术中相似的文章
[李义翠]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。