基于RGB-D大规模数据集的人体行为识别算法研究

发布时间：2019-05-24 15:31

【摘要】：进入21世纪,随着信息技术的发展和人类生活的不断智能化,计算机视觉正日益影响到人们生活的各个方面,而人体行为识别及分析,因其广泛的应用前景和实用价值,近几年来一直都是计算机视觉方向的研究热点。人体行为识别,即对原始的视频图像序列进行分析,提取相关的行为特征信息,最后对这些信息进行解释以实现对人体行为的识别和学习。尽管计算机技术和图像处理技术的快速发展极大地推动了行为识别领域的研究,而且随着大数据技术的普及,算法的性能越来越多地依赖于数据集,然而如何选择有效行为特征,以及目前数据集存在的遮挡、背景单一和缺少大量样本数据等问题,使得基于大量数据的复杂自然场景下的人体行为识别技术仍然是一个非常具有挑战性的研究领域。彩色-深度(RGB-Depth,RGB-D)传感器能够同时提供彩色和深度图像,3D深度信息可以直接获取而不再需要额外的计算,这为深度信息在人体行为识别领域内的应用提供了很大的方便。人体行为的识别和分析是依托于行为数据集进行的,在行为识别研究的过程中,先后出现了多种数据集,目前存在的公共RGB-D行为数据集由于有限的行为类别,行为样本数量以及单一化的背景环境,很难用于基于大量数据的复杂自然场景下的行为识别。因此,本文建立了一个综合性RGB-D大规模行为数据集,以促进对复杂自然场景下人体行为识别的研究,同时,基于该综合性数据集应用了三种特征提取算法。本文的具体研究内容如下:第一,分析了人体行为识别的研究背景、意义和目的,从数据集、特征提取和分类器三个层面综述了人体行为识别的研究现状,对基于RGB-D的行为识别研究面临的问题进行了阐述,介绍了本文的主要研究内容和章节安排。第二,阐述了 RGB-D传感器的优点以及深度信息在人体行为识别中的重要性,对目前已经存在的一些RGB-D数据集进行了详细的介绍,并比较了它们的优缺点。第三,选择了五个常用典型的RGB-D数据集,通过对这五个数据集中的数据进行预处理、统计分析,最后整合成一个综合性的RGB-D大数据集,并且对RGB-D大规模数据集中的行为类别进行了重新标定,统一了数据存储格式。该部分主要对建立大规模数据集中的工作做了具体的描述,同时介绍了 RGB-D大规模数据集的数据信息、优势以及建立意义。第四,基于RGB-D大规模数据集提取了深度行为投影图(Depth Motion Maps,DMMs)、深度立方体相似性特征(Depth Cuboid Similarity Feature,DCSF)和曲率尺度空间(Curvature Space Scale,CSS)三种类型的特征。DMMs特征累计整个深度视频序列里两个连续帧投影图之间的绝对差(运动能量);DCSF描述了围绕时空兴趣点构造的尺度自适应3D深度立方体之间的相似性关系;CSS可以表示不同尺度水平下人体轮廓曲线的不变特征。三种特征提取算法分别在五个子数据集和综合性大数据集上进行实验,应用协作表示分类器(Collaborative Representation Classifier,CRC)对人体行为进行识别,通过对实验结果的比较和分析来验证所建立RGB-D大规模数据集的适用性和有效性。最后,对本文所做的全部工作进行整理总结,并对未来的研究方向进行展望。
[Abstract]:In the 21st century, with the development of information technology and the increasingly intelligent of human life, computer vision is increasingly affecting all aspects of people's life, and human behavior identification and analysis, because of its wide application prospect and practical value, In recent years, it has been a hot topic in computer vision. The human behavior recognition, that is, the original video image sequence is analyzed, relevant behavior characteristic information is extracted, and finally, the information is interpreted so as to realize the identification and learning of human behavior. Although the rapid development of computer technology and image processing technology has greatly promoted the research in the field of behavior recognition, and with the popularization of large-data technology, the performance of the algorithm is increasingly dependent on the data set, however, how to select the effective behavior feature, As well as the problems such as occlusion, background single and lack of large amount of sample data in the current data set, the human behavior recognition technology under the complex natural scene based on the large amount of data is still a very challenging research field. The color-depth (RGB-Depth, RGB-D) sensor can provide both color and depth images at the same time, and the 3D depth information can be directly acquired without additional calculation, which provides great convenience for the application of the depth information in the field of human behavior identification. The identification and analysis of human behavior is based on the behavior data set. In the course of the study of behavior recognition, a variety of data sets have been presented, and the existing common RGB-D behavior data sets are due to the limited behavior category, the number of behavior samples and the single background environment. It is difficult to use for behavior recognition in complex natural scenes based on a large amount of data. Therefore, a comprehensive RGB-D large-scale behavior data set is established to promote the research of human behavior recognition in complex natural scene, and three feature extraction algorithms are applied based on the comprehensive data set. The research contents of this paper are as follows: First, the research background, meaning and purpose of human behavior recognition are analyzed, the research status of human behavior recognition is summarized from three aspects of data set, feature extraction and classifier, and the problems facing the research of behavior recognition based on RGB-D are described. The main contents and chapters of this paper are introduced. Secondly, the advantages of RGB-D sensor and the importance of depth information in human behavior recognition are described, and some of the existing RGB-D data sets are described in detail, and their advantages and disadvantages are compared. thirdly, five typical RGB-D data sets are selected, the data in the five data sets are pre-processed, analyzed and finally integrated into a comprehensive RGB-D large data set, and the behavior categories in the RGB-D large-scale data set are re-calibrated, The data storage format is unified. This part mainly describes the establishment of large-scale data set, and introduces the data information, advantage and significance of the large-scale data set of RGB-D. Fourth, based on the RGB-D large-scale data set, three types of features of the depth behavior projection (DMM), the depth cube similarity feature (DCSF) and the curvature scale space (CSS) are extracted. The DMs feature accumulates the absolute difference (motion energy) between two consecutive frame projections in the entire depth video sequence; the DCSF describes the similarity relationship between the scale adaptive 3D depth cubes constructed around the space-time interest point structure; CSS can represent the invariant feature of the human profile curve at different scale levels. The three feature extraction algorithms are tested on the five sub-data sets and the comprehensive large data set, and the cooperative expression classifier (CRC) is used to identify the human behavior. The applicability and validity of the established RGB-D large-scale data set are verified by the comparison and analysis of the experimental results. Finally, the whole work done in this paper is summarized, and the future research direction is expected.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】