基于Kinect骨骼关节信息的视角不变步态识别方法

发布时间：2018-02-16 21:07

本文关键词： 生物特征识别步态识别 Kinect 步态数据库特征融合　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：步态识别是一种利用人走路姿势进行人物身份识别的新兴生物特征识别技术。步态识别不同于一般生物识别的最大特点在于它可以在远距离且不需要测试者配合的情况下完成识别过程。同时,步态识别作为一种行为生物特征识别方法,就如同人们的书写方式一样,自然流畅且难以模仿。基于这些传统生物特征所不具有的优势,步态识别在门禁系统、安全监控、人机交互、医疗诊断等领域具有广泛的应用前景和经济价值。随着近些年视频影像技术越来越发达,模式识别算法越来越智能,以及计算机硬件处理能力越来越强大,这些都让步态识别换发出更新鲜的活力,受到越来越多研究人员的关注。在现如今的步态识别领域,基本可以分为基于2D视频的方法和基于3D信息的方法。基于2D视频的方法配置简单,但易受光照,视角变化的影响。基于3D信息的方法虽然可以很大程度上减少视角变化所引起的影响,但使用多相机重构3D信息的方法需进行相机标定,步骤繁杂,应用效果较差,而使用深度传感器的方法虽可避免相机标定,但深度相机价格昂贵,市场推广性较差。近年来,随着Kinect的问世,深度传感器趋于商业化、廉价化。这种深度传感器不仅不易受光线影响,而且易于利用深度信息分割运动人体与背景信息,同时可以便捷高效的获取人体关节3D信息,从而为步态识别的应用开启了一扇新的大门。本文的主要创新和贡献主要在以下两个方面:(1)本文以第二代Kinect为工具,利用其易于提取人体骨骼关节3D位置信息和基于深度信息易于分割人体与背景的特点,建立了一个全新的基于Kinect的步态数据库。该数据库采集了 52个个体在6个固定视角和4个自定义视角下的1040条行走片段。每条行走片段中包含该个体每帧21个骨骼关节的3D位置信息和每帧对应的二值化人体剪影轮廓图。(2)本文利用Kinect所采集到的关节位置信息,提取了大腿长度等8个骨骼长度作为静态特征,同时提取了大腿关节等8个随时间变化的关节摆动角度作为动态特征,并且分别验证了这两种特征对视角变化的鲁棒性。针对动态特征,本文同时也讨论了周期提取,模板匹配时相似性度量等关键性问题。在此基础上,本文对这两种特征在匹配层进行了特征融合,并利用最近邻分类器进行分类识别。最后,本文分别对静态特征,动态特征以及融合特征在本数据库下考察了视角不变与视角变化时的识别效果,并将本文的方法与传统基于2D视频的方法,基于Kinect的方法做了对比,同时在另一个基于Kinect的数据库上验证了本文的算法。结果表明本文提出的方法对视角变化有较好的鲁棒性,识别效果较好,并且算法简单,实时性强。本文建立的步态数据库同时支持验证基于Kinect关节位置的算法以及基于2D剪影轮廓图的算法,并为两者相互对比提供了可能,展现了本数据库所具有的独特优势。
[Abstract]:Gait recognition is a new biometric recognition technique based on human walking posture. Gait recognition is different from general biometric recognition in that it can be used for long distance and does not require testers to cooperate. To complete the identification process. At the same time, Gait recognition, as a behavioral biometric recognition method, is natural and difficult to imitate, just like the way people write. Based on the advantages of these traditional biometrics, gait recognition is in the access control system, security monitoring, With the development of video image technology, pattern recognition algorithm is more and more intelligent, and the computer hardware processing ability is more and more powerful, with the wide application prospect and economic value in the fields of man-machine interaction, medical diagnosis and so on. All of this has given gait recognition a fresher life and attracted more and more attention from researchers. In the field of gait recognition today, It can be divided into 2D video based method and 3D information based method. 2D video based method is easy to configure, but it is easy to be illuminated. Although the method based on 3D information can reduce the influence caused by the change of view angle to a great extent, the method of using multi-camera to reconstruct 3D information needs camera calibration, the steps are complicated, and the application effect is poor. Although the depth sensor can avoid camera calibration, the depth sensor is expensive and less widely used. In recent years, with the advent of Kinect, the depth sensor tends to be commercialized. This kind of depth sensor is not only easy to be affected by light, but also easy to use depth information to segment moving human body and background information, and can obtain 3D information of human joint easily and efficiently. Thus, it opens a new door for the application of gait recognition. The main innovation and contribution of this paper are mainly in the following two aspects: 1) the second generation Kinect is used as a tool in this paper. It is easy to extract 3D position information of human skeletal joint and to segment human body and background based on depth information. A new gait database based on Kinect was established. The database collected 1, 040 walking fragments of 52 individuals with 6 fixed and 4 custom views. Each walking fragment contains 21 pieces per frame of the individual. The 3D position information of skeletal joints and the corresponding binary silhouette contour of each frame are used in this paper, which is based on the joint position information collected by Kinect. Eight bone lengths, such as thigh length, were extracted as static features, and eight joints swinging angles, such as thigh joints, were extracted as dynamic features. And the robustness of the two features to the change of view angle is verified respectively. For dynamic features, this paper also discusses the key problems such as period extraction, similarity measurement of template matching and so on. In this paper, the two features are fused in the matching layer, and the nearest neighbor classifier is used to classify the two features. The dynamic feature and fusion feature are studied in this database when the angle of view is invariant and the angle of view changes. The method of this paper is compared with the traditional method based on 2D video and the method based on Kinect. At the same time, the algorithm is verified on another database based on Kinect. The results show that the proposed method is robust to the change of view angle, and the recognition effect is good, and the algorithm is simple. The gait database established in this paper supports the algorithms based on Kinect joint position and 2D silhouette at the same time, and provides the possibility for the comparison between the two, which shows the unique advantages of this database.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】