不规范书写坐姿的多类特征融合识别方法研究
发布时间：2019-04-12 13:45
【摘要】：随着社会的发展,我国少年儿童升学压力越来越大,学习时间越来越长,近视发病率也随之越来越高。据专家分析表明长时间的不良书写坐姿是导致少年儿童近视和生长发育不良主要原因之一。良好的学习坐姿与他们的生长发育息息相关,及时地纠正和保持正确的坐姿对孩子的健康成长非常重要。纸质练习簿无法实现对学生书写坐姿的监督,预防方法主要靠老师、家长提醒,身体状况出现问题后通过长时间地佩戴耳挂式或脊椎式坐姿矫正器。文字书写教学系统为实时监督书写坐姿提供了技术条件。不规范书写坐姿检测作为该系统中交互结构方面的重要研究分支,为用户不良坐姿的自动提醒和矫正提供真实、可靠的坐姿数据,是其不可或缺的前置工作。目前,坐姿识别以单类特征为主,如基于用户轮廓的几何特征法给出一种有害人体姿势报警方案,基于肤色特征的坐姿识别方法等。单类特征识别不规范书写坐姿的主要缺点是识别率偏低,本文针对低龄用户的一些独特不规范书写坐姿提出多类特征融合识别方法。论文的主要工作及成果如下:第一、对坐姿检测过程中涉及到相关检测识别算法进行了详细地阐述,介绍了运动目标检测、肤色特征提取、SURF特征提取及多分类算法等算法的适应情况和优缺点,为本文后续工作提供理论支持。第二、从生理学角度分析了坐姿原理。针对少年儿童坐姿的特点,将用户书写坐姿分为正确、趴写、含笔、托腮、歪头、驼背等8种类别。通过对用户坐姿分析,应用肤色在YCbCr空间聚集在一片固定区域且在CbCr平面上投影为一个近似椭圆的特性,提取各类坐姿在不同亮度下的肤色特征,以肤色特征描述手、脸的空间位置关系;依据不同阈值进行坐姿的SURF特征提取,以SURF特征分布特点来表征坐姿特征信息。第三、针对单类特征识别不规范书写坐姿识别率偏低的现状,提出多类特征融合的不规范书写坐姿分类方法。经单类特征分类得到各类坐姿识别正确率,计算出同类坐姿异类特征的融合权值,即同类坐姿的异类特征权值之和为一,且与识别率成正比例关系。然后,同类坐姿异类特征加权融合,BP神经网络训练识别。第四、实现不规范坐姿的监测方案。基于同类坐姿异类特征加权融合算法,设计了坐姿检测仿真系统,以检验算法的可行性,并对单类特征坐姿识别方法进行了对比实验,以体现算法的优越性。经实验分析表明,该方法的不规范书写坐姿识别率比单类特征法有明显提高,可为文字书写教学系统的不良坐姿的自动提醒和矫正提供真实、可靠的坐姿数据,能够提高少年儿童的不良坐姿检测识别率,具有更好地实用性。
[Abstract]:With the development of society, the pressure of children entering school is increasing, the study time is longer and the incidence of myopia is higher and higher. According to the analysis of experts, long-term poor writing and sitting posture is one of the main causes of childhood myopia and growth dysplasia. Good study of sitting posture is closely related to their growth and development. It is very important to correct and maintain correct sitting posture in time for the healthy growth of children. The paper exercise book can not realize the supervision of students' writing sitting posture. The preventive methods mainly rely on teachers and parents to remind them that after physical problems, they wear ear-hanging or spine-type sitting posture correction devices for a long time. Writing teaching system provides technical conditions for real-time supervision of sitting posture. As an important research branch in the interactive structure of the system, non-standard writing sitting posture detection provides real and reliable sitting posture data for the automatic reminder and correction of the user's bad sitting posture, and is an indispensable pre-work of the system. At present, sitting pose recognition is mainly based on one kind of features, such as geometric feature method based on user profile gives a harmful human pose alarm scheme, sitting pose recognition method based on skin color feature and so on. The main disadvantage of one-class feature recognition is that the recognition rate is low. This paper proposes a multi-class feature fusion recognition method for some unique non-standard writing sitting posture of young users. The main work and achievements of this paper are as follows: firstly, the related detection and recognition algorithms are described in detail in the process of sitting and pose detection, and the moving target detection and skin color feature extraction are introduced. The adaptation, advantages and disadvantages of SURF feature extraction and multi-classification algorithm provide theoretical support for the follow-up work in this paper. Secondly, the principle of sitting posture is analyzed from the point of view of physiology. According to the characteristics of children's sitting posture, the user's writing sitting posture is divided into 8 categories, such as correct writing, writing down, writing with pen, supporting gills, crooked head, hunchback and so on. By analyzing the user's sitting posture, the skin color is gathered in a fixed area in the YCbCr space and projected on the CbCr plane as an approximate ellipse. The skin color features of various sitting posture under different luminance are extracted, and the hand is described by the skin color feature. The spatial position relationship of face; The SURF features of sitting posture were extracted according to different thresholds, and the feature information of sitting posture was represented by the distribution of SURF features. Thirdly, aiming at the low recognition rate of non-standard writing sitting posture in one-class feature recognition, a multi-class feature fusion method is proposed to classify the unstandardized writing sitting posture. The correct rate of seat pose recognition is obtained by one-class feature classification, and the fusion weights of different features of the same sitting posture are calculated, that is, the sum of the weights of different features of the same sitting posture is one, and it has a positive proportional relationship with the recognition rate. Then, the weighted fusion of different features of the same sitting posture and BP neural network training recognition. Fourth, the implementation of the non-standard sitting posture monitoring scheme. Based on the weighted fusion algorithm of the same sitting posture features, the simulation system of sitting posture detection is designed to verify the feasibility of the algorithm, and a comparative experiment is carried out to show the superiority of the algorithm. Experimental analysis shows that the recognition rate of non-standard writing sitting posture by this method is obviously higher than that of single-class feature method, and it can provide true and reliable sitting posture data for automatic reminder and correction of bad sitting posture in Chinese writing teaching system. It can improve the recognition rate of children's bad sitting and pose detection, and has better practicability.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
，
本文编号：2457071
[Abstract]:With the development of society, the pressure of children entering school is increasing, the study time is longer and the incidence of myopia is higher and higher. According to the analysis of experts, long-term poor writing and sitting posture is one of the main causes of childhood myopia and growth dysplasia. Good study of sitting posture is closely related to their growth and development. It is very important to correct and maintain correct sitting posture in time for the healthy growth of children. The paper exercise book can not realize the supervision of students' writing sitting posture. The preventive methods mainly rely on teachers and parents to remind them that after physical problems, they wear ear-hanging or spine-type sitting posture correction devices for a long time. Writing teaching system provides technical conditions for real-time supervision of sitting posture. As an important research branch in the interactive structure of the system, non-standard writing sitting posture detection provides real and reliable sitting posture data for the automatic reminder and correction of the user's bad sitting posture, and is an indispensable pre-work of the system. At present, sitting pose recognition is mainly based on one kind of features, such as geometric feature method based on user profile gives a harmful human pose alarm scheme, sitting pose recognition method based on skin color feature and so on. The main disadvantage of one-class feature recognition is that the recognition rate is low. This paper proposes a multi-class feature fusion recognition method for some unique non-standard writing sitting posture of young users. The main work and achievements of this paper are as follows: firstly, the related detection and recognition algorithms are described in detail in the process of sitting and pose detection, and the moving target detection and skin color feature extraction are introduced. The adaptation, advantages and disadvantages of SURF feature extraction and multi-classification algorithm provide theoretical support for the follow-up work in this paper. Secondly, the principle of sitting posture is analyzed from the point of view of physiology. According to the characteristics of children's sitting posture, the user's writing sitting posture is divided into 8 categories, such as correct writing, writing down, writing with pen, supporting gills, crooked head, hunchback and so on. By analyzing the user's sitting posture, the skin color is gathered in a fixed area in the YCbCr space and projected on the CbCr plane as an approximate ellipse. The skin color features of various sitting posture under different luminance are extracted, and the hand is described by the skin color feature. The spatial position relationship of face; The SURF features of sitting posture were extracted according to different thresholds, and the feature information of sitting posture was represented by the distribution of SURF features. Thirdly, aiming at the low recognition rate of non-standard writing sitting posture in one-class feature recognition, a multi-class feature fusion method is proposed to classify the unstandardized writing sitting posture. The correct rate of seat pose recognition is obtained by one-class feature classification, and the fusion weights of different features of the same sitting posture are calculated, that is, the sum of the weights of different features of the same sitting posture is one, and it has a positive proportional relationship with the recognition rate. Then, the weighted fusion of different features of the same sitting posture and BP neural network training recognition. Fourth, the implementation of the non-standard sitting posture monitoring scheme. Based on the weighted fusion algorithm of the same sitting posture features, the simulation system of sitting posture detection is designed to verify the feasibility of the algorithm, and a comparative experiment is carried out to show the superiority of the algorithm. Experimental analysis shows that the recognition rate of non-standard writing sitting posture by this method is obviously higher than that of single-class feature method, and it can provide true and reliable sitting posture data for automatic reminder and correction of bad sitting posture in Chinese writing teaching system. It can improve the recognition rate of children's bad sitting and pose detection, and has better practicability.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
，
本文编号：2457071
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2457071.html
