摘 要

人体行为识别是近年来计算机视觉研究中的研究热点也是难点，在智能监控、虚拟现实、运动分析等诸多领域均有广阔的应用前景。基于视频流的人体行为识别包括主要包括运动目标检测、特征提取和行为识别三个部分，其中特征提取是最为关键的一个环节。特征提取的目的是从图像的众多数据中找出最能代表人体动作的有效特征，特征提取的好坏直接影响到下一步分类器对行为识别的准确性。本文针对视频流运动人体行为识别的特征提取问题进行了研究，主要工作如下：

(1)利用图像预处理方法（灰度化、高斯滤波、中值滤波等方法）对原始视频流进行处理，采用基于混合高斯背景建模的背景减除法提取出运动人体的轮廓。为了得到更优质的轮廓，对轮廓进行形态学处理、空洞填补以及连接相邻连通域等前景后处理的方法。所谓连接相邻连通域：当背景减除法得到的轮廓断裂成若干个连通区域时，计算区域的中心点位置，在此基础上计算连通域之间的距离，当距离小于一定的阈值时，则判断为两个区域同属一个人体并连通这两个区域。

(2)使用最小二乘法拟合的运动人体轮廓的外接椭圆，使用椭圆的长短轴比和椭圆的倾斜角度来反映人体的运动剧烈程度和倾斜程度，并作为人体的运动特征信息。为了保证特征具有尺度不变性，还采用了轮廓质心的方差以及Hu矩作为人体的特征信息。

(3)将上述提取的特征信息构成特征向量，运用支持向量机分类器识别人体的行为。

关键词：特征提取 人体行为识别 视频流 人体轮廓

On feature extraction of human behavior recogonition in video streaming

Abstract

The analysis and recognition human behavior is a hot research topic in the field of computer vision and this is also a difficult problem. It has broad application prospects in many fields, such as intelligent monitoring, medicine, sports and virtual reality. Human behavior recognition based on video stream includes three parts, including moving object detection, feature extraction and behavior recognition, and feature extraction is the most important one. The purpose of feature extraction is to find out the most effective features of human motion from image data, and the feature extraction has a direct impact on the accuracy of the next classifier for behavior recognition. In this paper, the problem of feature extraction of human behavior recognition in video stream motion is studied as follow.

1) Image preprocessing method, including gray, Gaussian filtering and median filtering of the original video stream processing. Using Gaussian mixture background model based on background subtraction method to extract the contours of the human motion. In order to get a better contour, the method of morphological processing, cavity filling and the adjacent connected domain are obtained. The so-called connecting adjacent connected domain: when the background subtraction method is used to get the contour of the fracture into several connected regions to find all connected domain of the center position, to determine each two connected inter domain distance, when the distance is less than a certain threshold, judging for two connected domain belong to one body, with wide line connecting the two connected domain in the center point to connect the two connected domain.

2) Using least square method to fit the contour connected elliptic represent human movement, using the elliptical ratio of the major and minor axis and elliptic to reflect human motion intensity and degree of inclination, and as the characteristic information of the human body. In order to ensure that the features are invariant to scale, the variance of the centroid and the Hu moments are used as the characteristic information of the human body.

3）Using the extracted feature information to form a multi-dimensional feature vector. Use a SVM classifier to recognize human behavior.

Key Words: Feature Extraction; Human Action Recognition; Video Stream; Human Body Contour

目 录

摘要 I

ABSTRACT II

第一章 绪论 1

1.1 课题研究背景及意义 1

1.2 国内外研究现状与分析 3

1.3 本文研究的主要内容与论文安排 6

第二章 人体行为识别整体方案设计 7

2.1 系统设计目标 7

2.2 系统总体设计方案 7

2.3 系统硬件搭建 7

2.4 系统软件设计 9

2.4.1 视频采集模块 9

2.4.2 特征提取及识别模块 9

2.4.3 显示模块 9

2.3.4 数据管理模块 10

第三章 行为特征提取 11

3.1 图像预处理 11

3.1.1 灰度化 11

3.1.2 图像滤波 12

3.2 运动目标检测 13

3.2.1 混合高斯背景建模 14

3.2.2 前景后处理 15

3.2.3 前景后处理处理结果 18

3.3 行为特征提取 19

3.3.1 提取外部轮廓 19

3.3.2 连接相邻连通域 21

3.3.3 拟合轮廓外接椭圆 21

3.3.3 计算质心及质心方差 23

3.3.4 计算Hu不变矩 24

3.3.5人体行为特征的表示 25

3.4 实验结果与分析 25

3.4.1 特征信息 26

3.4.2 分类信息 30

3.4.3 识别结果与分析 31

第四章 视频流中的人体行为识别系统实现 32

4.1 软件工具介绍 32

4.2 数据管理模块设计 33

4.3 方案整体实现 33

4.3.1 系统实现框图 33

4.3.2 系统图形界面 36

第五章 本文总结与展望 42

5.1 本文总结 42

5.2 未来展望 42

参考文献 43

致谢 45

附录 46

第一章 绪论

1.1 课题研究背景及意义

相关图片展示：