基于视频的多目标运动人体行为识别

发布时间：2018-06-15 20:35

本文选题：四元数 + 动作分割　；参考：《东华大学》2017年硕士论文

【摘要】：视觉是人类观察和认识世界的重要途径,基于视频的运动人体行为分析的目的是在不需要人为干预或者尽量少加入人为干预的条件下,实现基于视频的运动人体检测、跟踪、识别人的独立动作、认知多人之间以及人与周围环境之间的交互行为等。基于视频的人体行为分析,在智能视频监控智能人机交互以及人工智能领域均有重大的应用价值。本文使用微软Kinect for Windows 2.0作为人体运动视频获取装置,结合人体三维旋转信息,以现如今基于视频的人体行为分析研究中部分重点难点为出发点,对人体运动特征的提取、连续动作分割以及人体行为识别进行了探讨研究,获得了如下几个方面的成果:(1)定义深度图像的帧间距离实现人体连续动作的分割与识别。针对现有的动作识别方法大多数是建立在假设待识别人体行为已经给定动作的开始帧和结束帧的基础上这一问题,本文定义了深度图像的帧间距离,提出基于视频帧间距离(DBKF)的运动人体行为分割算法,实现了简单方便的单一运动目标的连续动作分割。(2)基于关节四元数的多目标人体行为识别。本文基于四元数理论结合三维人体骨架提出人体关节四元数作为特征信息用于建立运动人体行为表示模型,先将多目标人体进行个体分离,再使用支持向量机(SVM)作为分类器对视频片段中的人体运动行为进行分类,实现了少特征高准确度的基于视频的多目标人体行为识别。(3)基于隐马尔可夫模型(HMM)的异常行为识别。本文将三维人体旋转信息进行拓展运用到异常行为识别中,使用关节角作为特征参数,结合隐马尔可夫链模型算法(HMM),实现了对视频中的连续运动人体行为进行异常识别,并与经典的团块特征进行对比,分析了不同特征信息对异常行为识别准确性的影响。最后,针对论文的研究内容,探讨了研究中的可优化的方面,并就此得出新的研究方向。
[Abstract]:Vision is an important way for human beings to observe and understand the world. The purpose of video-based human motion human behavior analysis is to realize video-based human motion detection and tracking without human intervention or as little human intervention as possible. Identify individual actions and recognize interactions between people and their surroundings. Human behavior analysis based on video has great application value in intelligent video surveillance intelligent human-computer interaction and artificial intelligence. In this paper, we use Microsoft Kinect for Windows 2.0 as human motion video acquisition device, combined with 3D rotation information of human body, taking some key and difficult points in the research of human behavior analysis based on video nowadays as the starting point to extract human motion features. Continuous motion segmentation and human body behavior recognition are studied. The following results are obtained: 1) Defining the distance between frames of depth image to realize the segmentation and recognition of human continuous action. In order to solve the problem that most of the existing motion recognition methods are based on the assumption that the human behavior has been given a given start and end frame, this paper defines the distance between frames of the depth image. A motion human behavior segmentation algorithm based on video frame distance (DBKF) is proposed, which realizes simple and convenient continuous motion segmentation of a single moving object. (2) Multi-target human behavior recognition based on joint quaternion is realized. Based on the quaternion theory and the three-dimensional human skeleton, this paper proposes that the quaternion of human joint is used as the characteristic information to establish the behavior representation model of moving human body. Firstly, the multi-target human body is separated from each other. Then the support vector machine (SVM) is used as the classifier to classify the human motion behavior in the video segment, which realizes the recognition of abnormal behavior based on the hidden Markov model (HMMM), which is based on the multi-target human behavior recognition based on video with less features and high accuracy. In this paper, the 3D rotation information of human body is extended to the recognition of abnormal behavior. The joint angle is used as the characteristic parameter and the hidden Markov chain model algorithm is used to realize the abnormal recognition of the continuous moving human behavior in video. Compared with the classical block features, the influence of different feature information on the accuracy of abnormal behavior identification is analyzed. Finally, according to the research content of the thesis, the optimization of the research is discussed, and a new research direction is obtained.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】