基于半监督协同训练和集成学习的人体动作识别研究

发布时间：2018-04-26 02:08

本文选题：人体动作识别 + 协同训练　；参考：《江苏大学》2017年硕士论文

【摘要】：随着科学技术的发展,人体动作识别逐渐成为人工智能和机器视觉领域一个重要的研究方向,具有广阔的发展前景和很强的实用价值。可应用于日常的视频监控、智能医疗、运动分析、人机智能交互等。同时,由于视频中场景的复杂性、动作类内变化,以及需要大量的有标注样本来训练泛化性能强的识别模型,这些都给人体动作识别的研究带来了挑战性。本文对人体动作识别的若干问题,特别是基于半监督的人体动作识别进行了较深入的研究。首先阐述了人体动作识别的选题背景与研究的目的和意义;其次概述了人体动作识别的关键技术,如关键帧提取技术、特征提取技术以及人体动作识别技术等。在目前人体动作识别的理论研究基础上,本文提出了基于混合式协同训练的人体动作识别方法和基于半监督集成学习的人体动作识别方法,并在以上两种算法的基础上设计开发人体动作识别原型系统,主要研究内容如下:1)提出了基于混合式协同训练的人体动作识别方法。针对目前人体动作视频中有标记数据不足的问题,提出了一种基于混合式协同训练的新型人体动作识别方法。该方法利用动作识别领域不同类型的识别方法来构建基分类器,并进行迭代的相互训练以提高泛化性能,可以降低标注成本并实现不同识别方法的优势互补,进而提高人体动作的识别精度。实验结果表明,本文所提出的算法可以有效地识别视频中的人体动作。2)提出了基于半监督集成学习的人体动作识别方法。针对协同训练类算法随着迭代次数的增加,基分类器的差异性会越来越小,以及迭代训练中产生的基分类器没有被充分利用的问题。提出了基于协同训练和集成学习相结合的人体动作识别方法。该方法对每个基分类器设置一个集合。将基分类器迭代训练过程中产生的中间分类器加入到各自的集合中,然后利用这个集合来选择伪标号数据。并定义了一个基于置信度的最大证据边缘函数来选择伪标号数据,最终利用该算法对人体动作进行识别。该方法能有效克服协同训练迭代过程中基分类器差异退化的问题,进一步提高人体动作识别的准确率。3)设计并实现了基于半监督协同训练和集成学习的人体动作识别的原型系统。采用面向对象语言C#和MATLAB进行编程,通过原型系统的运行测试,表明所提的方法可用于相应的人体动作识别,并且该原型系统界面友好、功能齐全、可维护性好。
[Abstract]:With the development of science and technology, human motion recognition has gradually become an important research direction in the field of artificial intelligence and machine vision, which has broad development prospects and strong practical value. It can be used in daily video surveillance, intelligent medical treatment, motion analysis, human-computer intelligent interaction and so on. At the same time, due to the complexity of the scene in the video, the changes in the action class, and the need for a large number of labeled samples to train the generalized recognition model, all these bring challenges to the research of human motion recognition. In this paper, some problems of human motion recognition, especially semi-supervised human motion recognition, are studied in depth. Firstly, the background of the topic selection and the purpose and significance of the research are introduced. Secondly, the key technologies of human motion recognition, such as key frame extraction, feature extraction and human motion recognition, are summarized. Based on the theoretical research of human motion recognition, this paper proposes a human motion recognition method based on hybrid cooperative training and a semi-supervised integrated learning method. On the basis of the above two algorithms, the prototype system of human motion recognition is designed and developed. The main research contents are as follows: (1) A human motion recognition method based on hybrid cooperative training is proposed. Aiming at the shortage of tagged data in human motion video, a new human motion recognition method based on hybrid cooperative training is proposed. This method uses different recognition methods in the field of action recognition to construct the base classifier, and carries out iterative training to improve generalization performance. It can reduce the labeling cost and realize the complementary advantages of different recognition methods. And then improve the recognition accuracy of human body action. Experimental results show that the proposed algorithm can effectively recognize human motion in video. 2) A human motion recognition method based on semi-supervised integrated learning is proposed. As the number of iterations increases, the differences of base classifiers become smaller and smaller, and the basis classifiers generated in iterative training are not fully utilized. A method of human motion recognition based on cooperative training and integrated learning is proposed. The method sets a collection for each base classifier. The intermediate classifier generated in the iterative training process of the base classifier is added to the respective set, and then the pseudo-label data is selected by using this set. A maximum evidence edge function based on confidence degree is defined to select pseudo-label data. Finally, the algorithm is used to identify human actions. This method can effectively overcome the problem of the difference degradation of the base classifier in the iterative process of cooperative training. The prototype system of human motion recognition based on semi-supervised cooperative training and integrated learning is designed and implemented. Programming with object oriented languages C # and MATLAB, the test results of the prototype system show that the proposed method can be used for human body action recognition, and the prototype system has friendly interface, complete functions and good maintainability.
【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】