【作者】 杨海涛；

【导师】 马婷；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2016， 硕士

【摘要】 随着时代的发展,姿态分析和识别技术也不断发展,其应用范围也不断扩大。在本文中,提出了一种新的方案运用于人体姿态分析和识别,与传统的姿态捕捉技术不同,基于可穿戴惯性传感器的人体动作分析与识别是模式识别的一个新兴领域,克服了传统基于视频的动作识别的诸多缺点和限制,具有更高的可操作性和实用性。其实质是通过固定在人体特定部位的惯性传感器采集人体的动作信息,通过无线传输模块传到PC机,进而对数据进行预处理、特征提取和选择、动作分类。在医学康复工程、体感游戏领域、影视作品创作、虚拟现实、专业姿态分析等许多方面,人体姿态分析和识别技术获得越来越广泛的应用于并产生巨大的价值。本文首先对人体行为做了系统分析,明确了本文基于惯性传感器的采集方案,并对采集行为数据格式做了细致的分析。在此基础上,分析了人体关节、人体模型、关节角以及动作角,并将姿态角应用到人体行为分析中,明确了髋关节姿态角作为识别的原始数据。进而明确了人体结构和多刚体模型,并将其做了简单改进,确定了15个部位作为传感器佩戴位置,利用基于惯性传感器的动作捕捉系统对动作数据进行了采集,并建立基于惯性传感器的数据库。其次,通过对生活中人们的各种姿态的归纳、总结、分析以及深入研究,进而规范了数据库的数据采集和采集条件。然后,借用算法这种手段,对运动特征和数据预处理进行了识别和提取。当然还要想到外部噪音的影响,针对原始数据进行预处理操作,去除掉和正常值相比有偏离,与巴特沃斯低通滤波是考虑采取行动的人通提取特征点的多维姿态角数据。最后用基于动态时间规整方法进行了动作识别,给出了于惯性传感器的各动作的识别率,并与基于统计分析算法隐马尔科夫模型识别方法作对比进行对比分析。本文主要贡献是在三方面进行了创新。第一,建立基于惯性传感器动作数据库,并提出了对动作数据预处理的方法和流程;第二,提出了以髋关节位置的姿态角主要识别特征,明确了关键特征帧提取算法,得到了符合状态空间的特征序列;第三,提出了基于动态时间规整识别算法,达到了利用多传感器对人体姿态识别的目的,同时使惯性传感器对存在于空间中的多样姿态识别率有所提高。更多还原

【Abstract】 With the development of the times,motion analysis and recognition technology is developing continuously,and its application scope is also expanding.In this paper,a new scheme is proposed for human motion analysis and recognition.Unlike the traditional motion capture technique,human motion analysis and recognition based on inertial sensors is an emerging field of pattern recognition.It overcomes the disadvantages of traditional video action recognition of the many shortcomings and limitations,with a higher operability and practicality.The essence is that the movement information of the human body is collected by the inertial sensor fixed in the specific part of the human body and transmitted to the PC through the wireless transmission module,and then the data is pre-processed,the feature extracted and selected and the action classified.Human behavioral analysis and recognition technology have been widely applied and produced great value in many fields,such as medical rehabilitation engineering,somatosensory game field,film and television works creation,virtual reality,professional sports analysis and so on.In this paper,the human motion is analyzed,and the acquisition scheme based on the inertial sensor is defined,and the format of the collected motion data is analyzed in detail.On the basis of the analysis,the human joint,the human body model,the joint angle and the attitude angle are analyzed,and the attitude angle is applied to the human body motion analysis,and the hip joint angle is identified as the original data.Then the human body structure and rigid body model were defined,and 15 parts were selected as sensors.The movement data were acquired by inertial sensor-based motion capture system,and a database based on inertial sensor has been built.Secondly,we did a detailed research on the action level,summed up a variety of basic movements contained in the daily human movement,and standardize the data collection and collection conditions.After that,the related algorithms of data preprocessing and motion feature extraction and recognition are studied.Considering the noise of the system,the original data are preprocessed to eliminate the large deviation from the normal value,and the Butterworth low-pass filtering is taken into account to take the low-pass of the human action to extract the feature points for the multidimensional attitude angle data.Finally,the motion recognition based on the dynamic time regularization method is given,and the recognition rates of each action of the inertial sensor are given,and compared with the classical algorithm HMM identification method.The main contribution of this paper is innovation in three aspects.Firstly,a database based on the inertial sensor motion isestablished,and a method and a process for preprocessing the motion data are proposed.Secondly,the main recognition feature of the attitude angle of the hip position is proposed,and the feature extraction algorithm of the key feature frame is defined.And the feature sequence of state space is proposed.Thirdly,the identification of inertial sensors is realized based on the dynamic time regularization recognition algorithm and the recognition rate of inertial sensors based on spatial state motion are improved.更多还原