基于卷积神经网络的人体姿态识别研究

发布时间：2018-04-09 02:04

本文选题：人体姿态识别　切入点：深度图像　出处：《成都理工大学》2017年硕士论文

【摘要】：人体姿态识别是计算机视觉领域中最具挑战性的研究方向,被广泛应用于人机交互、智能监控、虚拟现实、人体行为分析等领域。随着微软公司推出Kinect深度传感设备之后,基于深度图像的姿态识别技术得到了越来越多的学者关注和研究。但是,在研究中由于图像受光照、阴影外界变化以及人体姿态类别的限制、特征选择的难度、机器学习的局限性、识别算法的性能等因素的影响,人体姿态识别研究进展缓慢,姿态识别的方法成为近些年的研究热点。目前,人体姿态识别采用的方法大多是基于Kinect的骨架信息和深度图像信息,通过对图像分割实现人体前景提取,再利用HOG、HIFT或骨架关键坐标对人体姿态特征进行提取和降维,最后对模型进行训练和分类。虽然这些方法在识别精确度和准确性方面效果比较好,但是由于人体复杂度较高,通过骨骼关键点和深度图像信息提取的特征往往维度也比较高,对深度图像进行特征提取前大多还需要对图像进行一些预处理,导致特征提取难度大和识别效率不高。因此,本论文提出了一种基于卷积神经网络的人体姿态识别研究方法。主要研究内容和工作任务是:(1)概述了深度图像的成像原理和获取方法,通过Kinect拍摄不同体型的人在不同背景、光照条件下的图片,构建了包含十个姿势的训练样本以及测试样本,以保证数据库的多样性。(2)介绍了人工神经网络和卷积神经网络的相关理论。重点介绍了经典的前馈神经网络的结构及反向传播算法,以典型的卷积神经网络结构为例,详细阐述了网络结构的组成和训练方式以及卷积神经网络相关知识如激活函数、损失函数、梯度计算、softmax回归分类技术等。(3)针对卷积神经网络具有局部感知、权值共享、网络模型的复杂度低、图像可直接作为网络的输入以及对平移、比例缩放、倾斜等变形网络结构具有高度不变性等优点,本文提出了基于卷积神经网络的人体姿态识别研究,并从网络层数、滤波器个数、激活函数等方面优化网络结构,从而进行人体姿态的分类识别。
[Abstract]:Human posture recognition is the most challenging research field in the field of computer vision. It is widely used in human-computer interaction, intelligent monitoring, virtual reality, human behavior analysis and other fields.With the introduction of Kinect depth sensing device by Microsoft, the attitude recognition technology based on depth image has been paid more and more attention and research.However, due to the influence of illumination, shading outside changes and human pose types, the difficulty of feature selection, the limitation of machine learning, and the performance of recognition algorithm, the research progress of human pose recognition is slow.Attitude recognition has become a hot topic in recent years.At present, most of the human attitude recognition methods are based on the skeleton information and depth image information of Kinect. The human body foreground is extracted by image segmentation, and then the human pose feature is extracted and dimensionally reduced by using hog HIFT or skeleton key coordinates.Finally, the model is trained and classified.Although these methods have good recognition accuracy and accuracy, because of the high complexity of human body, the feature dimensions extracted from the key points and depth images of bone are often higher.Most of the depth images need to be preprocessed before feature extraction, which leads to the difficulty of feature extraction and the low efficiency of recognition.Therefore, this paper proposes a research method of human body attitude recognition based on convolution neural network.The main research content and work task is: (1) the imaging principle and acquisition method of depth image are summarized. The images of people of different body types under different background and illumination conditions are taken by Kinect.The training samples including ten postures and test samples are constructed to ensure the diversity of database. (2) the related theories of artificial neural network and convolutional neural network are introduced.The structure and back propagation algorithm of classical feedforward neural network are introduced in detail. Taking the typical convolutional neural network structure as an example, the composition and training mode of the network structure and the related knowledge of convolutional neural network such as activation function are described in detail.Loss function, gradient calculation and Softmax regression classification, etc.) the convolutional neural network has local perception, weight sharing, low complexity of the network model, and image can be directly used as the input of the network as well as translation and scaling.In this paper, the human body attitude recognition based on convolution neural network is proposed, and the network structure is optimized in terms of the number of network layers, the number of filters, the activation function, etc.Thus, the classification and recognition of human posture are carried out.
【学位授予单位】：成都理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】