论文总字数：30452字

摘 要

近年来随着服务型智能机器人技术的快速发展，如何使服务机器人能够高效地对人们日常行为识别成为了一个研究重点。只有正确地识别人们的行为，才可以使机器人或其它智能系统正确地为人提供服务或者协助人的工作。本文针对智能机器人需求，研究了基于RGB-D传感器的行为识别技术及其应用。

RGB-D传感器，深度信息的引入，为行为识别提供了更丰富准确的信息。本文使用了微软公司的Kinect传感器及SDK开发套件，提出了一种基于GMM聚类和DTW的行为识别算法。该方法通过人体骨架追踪，提取左右肩、双肘以及双手腕等结构组成特征向量，采用高斯混合模型（GMM）对行为时-空片段进行聚类，并研究了DTW和fastDTW两种识别算法，详细测试分析了两种算法的识别率和计算效率。在此基础上，以面向人机交互的交通手势识别和日常行为识别作为两类典型具体应用，开发了基于Kinect的行为识别软件，在自建的小规模样本上训练并测试了该方法具有较高的准确率。

本论文将先介绍整个系统的组织结构框架，再分块详细介绍算法的具体实现。随后以数据表格的形式，展示该算法进行行为识别的效果，并进行了分析。最后提出了这个系统的应用与发展以及改进。

关键词：行为识别 Kinect GMM聚类 DTW算法

THE RESEARCH AND APPLICATION ON DAILY LIVING BEHAVIOR RECOGNITION BASED ON DEPTH VISION

Abstract

In recent years, with the rapid development of intelligent service robot technology, how to efficiently identify people's daily behavior has become a research focus on service robots. Only by correctly identifying people's behavior, it can make a robot or other intelligent systems to properly provide services or assistance for people who work. In this paper, aiming at the demand for smart robots, we conduct research on behavior recognition technology based on RGB-D sensor and its application.

With the depth information from RGB-D sensor introduced, the behavior recognition is provided with richer and more accurate information. This paper used Microsoft's Kinect sensor and SDK development kit, proposed activity recognition algorithm based on GMM clustering and DTW. In this method, feature vectors are composed of structures of shoulders, elbows and wrists extracted by human skeleton tracking, and we used a Gaussian mixture model (GMM) of behaviors’ space-time fragments’ clustering, and studied DTW and fastDTW identification algorithms, and make detailed test and analyzes the recognition rate and the computational efficiency of the two algorithms. On this basis, traffic gesture recognition for human-computer interaction and daily life behavior recognition as two typical applications, we developed based behavior recognition software based on Kinect, and tested the method on a small self-built database and concluded with high accuracy rate.

The paper will start with the structure of the system, and demonstrate the algorithms in detail. Then a table of the results will be presented, showing the final results of activity recognition and make some analysis. Finally, we will talk about its application, further development and improvement.

Key words: Activity recognition, Kinect, GMM clustering, DTW algorithm

目录

摘要 2

Abstract 2

目录 3

第1章 绪论 4

1.1 背景与意义 4

1.2 相关技术研究进展 5

1.2.1 Kinect传感器简介 5

1.2.2 人体识别技术 6

1.2.3 行为识别研究现况 7

1.3 论文组织与结构 9

第2章 行为识别系统总体设计 10

2.1 设计需求分析 10

2.2 系统原理及总体架构 10

2.3 用户界面设计与实现 11

第3章 特征提取与预处理 14

3.1特征提取 14

3.1.1 骨骼跟踪 14

3.1.2行为特征选取 15

2.2 行为片段聚类 16

2.2.1 聚类方法概述 16

3.2.2 GMM模型 17

3.3 特征处理 19

3.3.1 特征处理具体方法 19

3.3.2 特征处理结果与分析 20

3.4 小结 22

第4章 基于DTW的行为识别 23

4.1 算法流程 23

4.2 DTW算法 23

4.2.1 传统的DTW算法 23

4.2.2 fastDTW算法 26

4.2.3 DTW算法速度测试 28

4.2.4 匹配过程 29

4.3 小结 30

第5章 实验结果与分析 31

5.1 行为识别算法测试 31

5.1.1 交通手势识别测试 31

5.1.2 日常行为识别测试 34

5.2 行为识别系统应用 36

5.2.1 骨骼跟踪模块 37

5.2.2 交互按钮模块 38

5.2.3 算法脚本模块 39

5.2.4 应用展示 40

第6章 总结与展望 40

参考文献 41

致谢 44

剩余内容已隐藏，请支付后下载全文，论文总字数：30452字