基于深度信息的行为表示与识别

发布时间：2018-05-24 19:18

本文选题：行为识别 + 深度信息　；参考：《中国民航大学》2017年硕士论文

【摘要】：近些年来,计算机视觉技术发展迅速。作为计算机视觉范畴的重要研究课题,行为识别具有重要的科学意义。基于RGB数据图像的人体行为识别容易受到光照强弱、阴影、以及外界环境的改变等因素的影响。并且,对于在二维平面没有明显变化,而在深度方向有较大变化的动作识别会比较困难。另外,目前已经提出的基于深度信息的行为识别,有的利用了非零像素计算深度差值,还有的利用了深度方向的变化。本文提出一种基于点云的动作描述方法——运动历史点云(MHPC)。运动历史图像(MHI)将动作视频压缩为一个二维图像。类似地,MHPC将动作视频压缩为一个包含深度信息的三维点云。点云的(x,y,z)通道用来记录人体动作过程中前景点的空间位置改变,颜色通道用来记录时序变化。这种方法有以下两个特点:1、MHPC在压缩数据的同时,保留了动作过程中的时序和空间变化信息,并且由于包含了深度变化信息,因此,在MHPC的基础上能够提取出丰富的特征以支持后续的动作识别;2、对于后续的特征提取和动作识别过程而言,MHPC比原始视频数据更方便处理,因此MHPC可以作为一种规范输入和强化动作特征的预处理手段。作为MHPC的一个应用实例,首先,用Harris3D从生成的运动历史点云中提取特征点,结合点特征直方图FPFH对特征点进行描述;然后,把特征描述子用K-means算法聚类生成单词包;最后,输入多类分类支持向量机进行动作的识别。并与相近方法进行了比较,实验结果证明了结合运动历史点云的人体行为识别方法的合理性和有效性。
[Abstract]:In recent years, computer vision technology has developed rapidly. As an important research subject of computer vision category, behavior recognition has important scientific significance. Human behavior recognition based on RGB data images is easily affected by the intensity of illumination, shadow, and the change of external environment. In addition, it is difficult to recognize the motion in the two-dimensional plane, but in the depth direction. In addition, the behavior recognition based on depth information has been proposed. Some of them use non-zero pixels to calculate the depth difference, and others make use of the change of depth direction. An action description method based on point cloud is presented in this paper. The motion history image MHI) compresses the motion video into a two-dimensional image. Similarly, the MHPC compresses the action video into a three-dimensional point cloud containing depth information. The point cloud channel is used to record the spatial change of the spot in front of the human body, and the color channel is used to record the temporal change. This method has the following two characteristics: 1: 1 MHPC compresses the data while retaining the temporal and spatial variation information in the course of action, and because it contains the depth change information, therefore, On the basis of MHPC, it can extract rich features to support subsequent motion recognition. It is more convenient to process the following feature extraction and motion recognition than the original video data. Therefore, MHPC can be used as a preprocessing method to standardize input and enhance action characteristics. As an application example of MHPC, firstly, feature points are extracted from the generated moving historical point cloud by Harris3D, and feature points are described by combining feature histogram FPFH. Then, feature descriptors are clustered by K-means algorithm to generate word packets. Input multi-class classification support vector machine to recognize the action. Compared with similar methods, the experimental results show the rationality and effectiveness of the human behavior recognition method combined with the point cloud of motion history.
【学位授予单位】：中国民航大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】