基于多传感器的人体姿态识别系统
发布时间：2019-05-14 23:03
【摘要】：人机交互系统就是通过人、机器与环境之间交换不同的信息、并进行理解与反馈的系统,如何更有效的实现人机沟通、理解人类成为未来人机交互的主要研究热点和方向,人体姿态识别技术作为其核心技术也受到越来越多的关注。目前主流研究方向有两个部分,一是专注于人体细节动作捕捉、虚拟现实等技术。另一类为专注于日常应用及健康看护的人体姿态识别技术,其目的在于理解人体的典型行为,从而完成人机交互。根据数据获得方式该技术主要分为两大类,基于图像的识别技术与基于传感器的识别技术。基于图像的识别是非接触式,适用于固定区域中的人体姿态识别,基于传感器的识别可以跟随目标进行数据采集,不受地域限制且基于当前的可穿戴技术,其传感器不会对接触处的节点运动产生影响。因此本文将设计应用于日常情景、基于传感器的人体姿态识别系统。识别的姿态包括“行走”、“跑步”、“原地剧烈运动”、“站立”、“倚坐”、“躺卧”、“摔倒”和“正常过渡状态”。首先本文设计了一种便携式的数据采集终端。基于基本的数据采集以及便携性要求,终端共分为四个单元:传感器单元、控制单元、存储单元和电源管理单元。终端采集数据稳定,待机时间长,为人体姿态识别系统原始数据来源。其次本次研究为了获得较纯净的数据,将对原始数据进行进一步的预处理及数据特征提取。预处理包括滤噪、定标以及去除重力加速度干扰。特征提取包括时域特征提取与频域特征提取。具体为样本帧的均值、标准差、中位绝对偏差、四分位距、最大值、最小值、平方和均值、伯格阶数为4的AR模型系数、各轴序列相关性、信息熵特征、峰值频点、次峰值频点、峰值带宽、加权平均频点、突出频点数量特征。经过不同样本间的比对,评估特征的可用性。最后构建了基于上述特征的人体姿态识别算法。首先将“行走”、“跑步”、“原地剧烈运动”定义为运动状态,将“站立”、“倚坐”、“躺卧”定义为静止状态,将“摔倒”和“正常过渡状态”定义为过渡状态。每种状态基于其大类特征的差异而采用支持向量机以及决策树等分类方式进行识分类。最后构成了较为成熟的人体姿态识别算法架构。
[Abstract]:Human-computer interaction system is a system in which different information is exchanged between people, machines and the environment, and how to realize human-computer communication more effectively and understand human beings becomes the main research focus and direction of human-computer interaction in the future. As its core technology, human attitude recognition technology has received more and more attention. At present, there are two main research directions, one is to focus on human detail motion capture, virtual reality and other technologies. The other kind of human posture recognition technology, which focuses on daily application and health care, aims to understand the typical behavior of human body and complete human-computer interaction. According to the data acquisition mode, the technology is mainly divided into two categories: image-based recognition technology and sensor-based recognition technology. The recognition based on image is non-contact and is suitable for human attitude recognition in fixed area. The recognition based on sensor can follow the target for data acquisition, which is not limited by region and based on the current wearable technology. The sensor has no effect on the motion of the node at the contact. Therefore, this paper applies the design to the daily situation, the human body attitude recognition system based on sensor. Identified postures include "walking", "running", "strenuous exercise in situ", "standing", "sitting", "lying", "falling" and "normal transition". First of all, a portable data acquisition terminal is designed in this paper. Based on the basic data acquisition and portability requirements, the terminal is divided into four units: sensor unit, control unit, storage unit and power management unit. The terminal acquisition data is stable and the standby time is long, which is the original data source of human attitude recognition system. Secondly, in order to obtain purer data, the original data will be further preprocessed and data feature extraction will be carried out. Preprocessing includes noise filtering, calibration and removal of gravity acceleration interference. Feature extraction includes time domain feature extraction and frequency domain feature extraction. Specifically, the mean value, standard deviation, median absolute deviation, quartile distance, maximum value, minimum value, square sum mean, AR model coefficient with Berg order 4, correlation of each axis sequence, information entropy characteristics, peak frequency point, Secondary peak frequency point, peak bandwidth, weighted average frequency point, highlight the number of frequency points. After comparison between different samples, the availability of features is evaluated. Finally, a human attitude recognition algorithm based on the above features is constructed. First of all, "walking", "running" and "strenuous exercise in situ" are defined as exercise state, "standing", "leaning" and "lying" are defined as static state, and "falling" and "normal transition state" are defined as transition state. Each state is classified by support vector machine (SVM) and decision tree based on the difference of its large class features. Finally, a more mature human attitude recognition algorithm architecture is formed.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP212.9
本文编号：2477117
[Abstract]:Human-computer interaction system is a system in which different information is exchanged between people, machines and the environment, and how to realize human-computer communication more effectively and understand human beings becomes the main research focus and direction of human-computer interaction in the future. As its core technology, human attitude recognition technology has received more and more attention. At present, there are two main research directions, one is to focus on human detail motion capture, virtual reality and other technologies. The other kind of human posture recognition technology, which focuses on daily application and health care, aims to understand the typical behavior of human body and complete human-computer interaction. According to the data acquisition mode, the technology is mainly divided into two categories: image-based recognition technology and sensor-based recognition technology. The recognition based on image is non-contact and is suitable for human attitude recognition in fixed area. The recognition based on sensor can follow the target for data acquisition, which is not limited by region and based on the current wearable technology. The sensor has no effect on the motion of the node at the contact. Therefore, this paper applies the design to the daily situation, the human body attitude recognition system based on sensor. Identified postures include "walking", "running", "strenuous exercise in situ", "standing", "sitting", "lying", "falling" and "normal transition". First of all, a portable data acquisition terminal is designed in this paper. Based on the basic data acquisition and portability requirements, the terminal is divided into four units: sensor unit, control unit, storage unit and power management unit. The terminal acquisition data is stable and the standby time is long, which is the original data source of human attitude recognition system. Secondly, in order to obtain purer data, the original data will be further preprocessed and data feature extraction will be carried out. Preprocessing includes noise filtering, calibration and removal of gravity acceleration interference. Feature extraction includes time domain feature extraction and frequency domain feature extraction. Specifically, the mean value, standard deviation, median absolute deviation, quartile distance, maximum value, minimum value, square sum mean, AR model coefficient with Berg order 4, correlation of each axis sequence, information entropy characteristics, peak frequency point, Secondary peak frequency point, peak bandwidth, weighted average frequency point, highlight the number of frequency points. After comparison between different samples, the availability of features is evaluated. Finally, a human attitude recognition algorithm based on the above features is constructed. First of all, "walking", "running" and "strenuous exercise in situ" are defined as exercise state, "standing", "leaning" and "lying" are defined as static state, and "falling" and "normal transition state" are defined as transition state. Each state is classified by support vector machine (SVM) and decision tree based on the difference of its large class features. Finally, a more mature human attitude recognition algorithm architecture is formed.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP212.9
【引证文献】
相关硕士学位论文 前1条
1 任相臻;基于多运动传感器的动态手势识别设计与实现[D];河北工程大学;2018年
，本文编号：2477117
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2477117.html
