基于运动特征的人体行为识别方法研究

发布时间：2018-05-29 10:47

本文选题：人体行为识别 + 可穿戴式动作捕捉系统　；参考：《安庆师范大学》2017年硕士论文

【摘要】：人体行为识别已经成为人机交互和普适计算的关键研究领域,其目标是提供关于用户行为的信息,允许计算机主动地帮助用户完成任务。计算机视觉研究一直处于这一工作的前沿。随着传感器技术的发展,研究人员逐渐开始使用惯性传感器(如加速度计或陀螺仪)进行行为识别研究。因此,基于无线传感器的人体行为识别方法正成为研究热点。特征提取是人体行为识别的核心问题之一,提取最有效的特征是提高人体行为识别精度的重要途径。现有的基于传感器的人体行为识别方法,在特征提取阶段往往采用诸如加速度或角速度数据的均值,方差,峰度等离散数据特征。一方面,这些离散型特征并不能反映人体运动的连续性,另一方面,这些通过手工选择的特征依赖于先验知识。因此,本文主要针对以上不足,采用基于传感器的可穿戴式动作捕捉系统,研究适用于人体行为识别的运动特征。其一,人体运动过程中,肢体的运动是连续的,而对应的动作捕捉数据是离散的。为了更好地分析人体日常行为的连续性与周期性,本文在第二章中提出了一种基于函数型数据分析的人体行为识别方法。首先,利用函数型数据分析方法,将可穿戴式动作捕捉系统采集的人体周期行为数据函数化,通过函数准确地定义数据的连续性与周期性。然后,从动作捕捉数据的函数中分别提取周期数据特征与函数型特征,并结合支持向量机方法,有效地识别了多类别的人体日常行为。其二,深度神经网络能够自动地学习低层次到高层次的分布式特征,并以此自学习特征代替传统方法中手工提取的特征。本文在第三章中定义了短时行为的概念,并基于短时行为提出了一种基于深度学习的人体行为识别方法。首先,利用滑动窗口分割的方法,构建了一个包含不同人体短时行为模式的过完备模式库。然后,基于此过完备模式库,使用卷积神经网络实现短时行为特征的自动学习,并最终实现对短时日常行为的分类识别。
[Abstract]:Human behavior recognition has become a key research field in human-computer interaction and pervasive computing. Its goal is to provide information about user behavior and to allow computers to actively help users complete tasks. Computer vision research has been at the forefront of this work. With the development of sensor technology, researchers began to use inertial sensors (such as accelerometers or gyroscopes) to identify behavior. Therefore, human behavior recognition based on wireless sensor is becoming a hot research topic. Feature extraction is one of the core problems in human behavior recognition. The most effective feature extraction is an important way to improve the accuracy of human behavior recognition. The existing sensor-based human behavior recognition methods often use discrete data features such as the mean variance kurtosis of acceleration or angular velocity data in feature extraction stage. On the one hand, these discrete features can not reflect the continuity of human motion, on the other hand, these features selected by hand depend on prior knowledge. Therefore, in this paper, a wearable motion capture system based on sensors is used to study the motion features of human behavior recognition. First, the body motion is continuous, and the corresponding motion capture data is discrete. In order to better analyze the continuity and periodicity of human daily behavior, a method of human behavior recognition based on functional data analysis is proposed in the second chapter. Firstly, the data of human periodic behavior collected by wearable motion capture system are functioned by the method of functional data analysis, and the continuity and periodicity of the data are accurately defined by the function. Then, the periodic data features and the functional features are extracted from the function of motion capture data, and the support vector machine (SVM) method is used to identify the human daily behavior effectively. Secondly, the deep neural network can automatically learn the distributed features from the lower level to the high level, and use the self-learning features to replace the features extracted by hand in the traditional methods. In the third chapter, the concept of short-term behavior is defined, and a human behavior recognition method based on deep learning is proposed based on short-term behavior. Firstly, an overcomplete pattern library containing different human short term behavior patterns is constructed by using sliding window segmentation method. Then, based on the over-complete pattern library, the convolution neural network is used to realize the automatic learning of short-time behavior features, and finally to realize the classification and recognition of short-time daily behavior.
【学位授予单位】：安庆师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP212.9;TP391.4

【参考文献】