基于深度图像数据的人体动作识别

发布时间：2018-04-26 09:31

本文选题：人体动作识别 + 深度图像　；参考：《郑州大学》2017年硕士论文

【摘要】：人体动作和行为识别在计算机视觉和模式识别等多个交叉领域一直以来都是一个热门的研究课题。它有广阔的应用前景如人机交互、智能家居、智能监控。早期的人体动作识别研究主要针对彩色可见光摄像机录制的视频数据,在该数据上提出了许多经典的算法。但是由于传统数据本身的局限性,这些算法容易受到光照变化等因素的影响。随着传感技术的发展,廉价的深度摄像机比如Kinect的出现,为研究者解决人体动作识别中的难题提供了新的思路。深度数据不仅不容易受光照变化等因素的影响,还能额外提供场景的三维信息,这使得越来越多的学者转向基于深度数据的动作识别算法的研究。本论文针对从深度图像数据中识别人体动作进行研究,实现了多种有效的人体动作识别算法,具体内容如下:第一,在深度图像的基础上,提出了一种动作特征描述方法。首先利用自适应深度运动图能量来划分时间动作序列,通过分析不同子时间段动作的姿态来获得运动能量模型(MEM),并利用局部二值模式(LBP)对其进行编码,得到MEM-LBP描述符来表示动作特征,经主成分分析(PCA)降维后,送入7)2范数协同表示分类器中进行识别。在MSR Action3D和MSR Gesture3D数据库上进行的测试分析与对比表明了该算法的有效性。第二,引入特征融合方法,在MEM的基础上,提取梯度方向直方图描述子(HOG)得到新的特征描述符,将其与MEM-LBP特征描述符进行特征层融合和决策层融合,并引入核极限学习机(KELM)对动作分类。在数据库上的测试实验结果表明两种融合方法都在一定程度上提升了识别效果,并且决策层融合的效果相对更好。
[Abstract]:Human motion and behavior recognition has been a hot research topic in many intersecting fields such as computer vision and pattern recognition. It has broad application prospects such as human-computer interaction, intelligent home, intelligent monitoring. The early research on human motion recognition mainly focused on video data recorded by color visible light camera, and put forward many classical algorithms on this data. However, due to the limitation of traditional data, these algorithms are easily affected by illumination changes and other factors. With the development of sensing technology, the emergence of cheap depth cameras such as Kinect provides a new way for researchers to solve the problem of human motion recognition. Depth data is not only not easily affected by illumination and other factors, but also can provide 3D information of scene. As a result, more and more scholars turn to the research of motion recognition algorithm based on depth data. In this paper, the recognition of human motion from depth image data is studied, and a variety of effective human motion recognition algorithms are implemented. The main contents are as follows: first, based on the depth image, an action feature description method is proposed. Firstly, the adaptive depth motion map energy is used to divide the time series of actions, and the motion energy model is obtained by analyzing the posture of the actions in different sub-time periods, and the local binary mode is used to encode the motion energy model. The MEM-LBP descriptor is obtained to represent the action feature. After the principal component analysis (PCA) is used to reduce the dimensionality, the MEM-LBP descriptor is sent to the 7 ~ (2) norm cooperative representation classifier for recognition. The test results on MSR Action3D and MSR Gesture3D database show that the algorithm is effective. Secondly, the feature fusion method is introduced, based on MEM, the gradient histogram descriptor is extracted to obtain the new feature descriptor, and the new feature descriptor is fused with the MEM-LBP feature descriptor at the feature level and decision level. And the nuclear extreme learning machine (KELM) is introduced to classify the action. The experimental results on the database show that the two fusion methods improve the recognition effect to a certain extent, and the effect of decision level fusion is better.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】