基于3D骨架片段表示度量及流形分割的人体动作识别

发布时间：2018-03-19 11:39

本文选题：人体动作识别　切入点：动作(片段)表示度量　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,由于人体动作识别在包括视频监控、人机交互娱乐、个人护理以及社会服务等方面的广泛应用,人体动作识别已成为当下计算机视觉领域最受重视的研究方向之一。人体动作识别的目标是让机器从一段视频中自动分析和识别所包含的动作。然而,动作视频中所固有的背景复杂、人体遮挡以及摄像机移动等问题,使得人体动作识别任务具有极大的挑战性。传统的研究工作专注于从二维视频帧图像中提取图像的局部时空特征,近年来,由于深度摄像机如微软体感摄像机Kinect以及与之对应的人体3D骨架坐标追踪算法的出现,促进了基于深度图像和人体3D骨架信息的一系列新的人体动作识别方法的发展。与二维视频帧图像相比,深度图像能够反应纯粹的几何和形状线索,且对照明条件的变化不敏感,因而基于深度图像的识别方法在实际应用中更具鲁棒性。与基于图像(RGB及深度图像)特征的人体动作识别方法不同,本文使用3D骨架信息表示人体动作:一方面,3D骨架信息在人体动作识别工作中的表现比其它基于视频帧图像特征提取的方法更好;另一方面,基于姿势序列的动作表示描述了人体动作的固有特征且符合人类自身对动作的学习认知过程。本文将人体动作局部基于3D骨架信息表示为人体姿势,整体上一系列姿势点看作人体动作流形,两者共同构成文章方法的主线。进而提出基于3D骨架片段表示及流形层次化分割和基于3D骨架片段距离度量及流形序列化分割两种人体动作识别方法。研究基于视频及其对应人体3D骨架坐标信息的动作识别问题,从视频的分割、分割片段的表示以及全局时间关系的处理三方面解决了人体动作识别问题。本文的主要工作和特色如下:(1)使用人体3D骨架信息表示人体动作,进而将人体动作(由人体姿势序列组成)当作流形进行处理;(2)基于3D骨架信息,提出主姿势特征和主要动作趋势特征表示人体动作(片段)并度量动作(片段)间的距离;(3)基于人体动作流形局部线性性度量实现人体动作的分解,利于表示及识别分类任务;(4)提出了两种人体动作识别方法,并在几个流行的人体动作数据集上进行实验,证明了所提方法的有效性。
[Abstract]:In recent years, human motion recognition has been widely used in video surveillance, human-computer interactive entertainment, personal care and social services. Human motion recognition has become one of the most important research areas in the field of computer vision. The goal of human motion recognition is to allow machines to automatically analyze and recognize the actions contained in a video. However, There are many problems in motion video, such as complicated background, occlusion of human body, moving of camera and so on. The traditional research work focuses on extracting local temporal and spatial features from two-dimensional video frame images. Because of the emergence of depth cameras such as Kinect, a Microsoft somatosensory camera, and the corresponding human 3D skeleton coordinate tracking algorithm, It promotes the development of a series of new human motion recognition methods based on depth images and human body 3D skeleton information. Compared with two-dimensional video frame images, depth images can reflect pure geometric and shape clues. Moreover, the method based on depth image is more robust in practical application because it is insensitive to the change of illumination condition. It is different from the method based on RGB and depth image. In this paper, 3D skeleton information is used to represent human action: on the one hand, 3D skeleton information is better than other methods based on video frame image feature extraction in human motion recognition; on the other hand, The action representation based on pose sequence describes the inherent characteristics of human actions and accords with the learning and cognition process of human actions. In this paper, the local human actions are represented as human posture based on 3D skeleton information. As a whole, a series of pose points are regarded as human action manifolds, Then, two human action recognition methods based on 3D skeleton segment representation and manifold hierarchical segmentation and 3D skeleton segment distance measurement and manifold serialization segmentation are proposed. The motion recognition problem of video and its corresponding 3D skeleton coordinate information of human body, The problem of human motion recognition is solved from three aspects: video segmentation, segment representation and processing of global temporal relationship. The main work and features of this paper are as follows: 1) Human body action is represented by 3D skeleton information. Then the human actions (composed of human pose sequences) are treated as manifolds based on 3D skeleton information. The main pose feature and the main action trend feature are proposed to represent the human action (segment) and measure the distance between the actions (segments). Based on the local linearity metric of human action manifold, the decomposition of human action is realized. In this paper, two methods of human motion recognition are proposed and tested on several popular human action data sets. The results show that the proposed method is effective.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】