基于视角归一化的步态识别研究

发布时间：2018-06-15 16:55

本文选题：步态识别 + 步态帧差熵图　；参考：《西安科技大学》2017年硕士论文

【摘要】：随着智能化程度的逐步深入,公共场合的安全形势日益严峻,如何确保公共安全成为人们关心和瞩目的焦点,而有效的身份识别技术则是确保公共安全的关键。生物特征以其安全、稳定、可靠等特点广泛应用于智能监控领域,较之于人脸、指纹等,步态以其非侵犯、远距离、难以隐藏等优势受到了大批研究者的关注。近年来,多视角下的步态识别问题一直是步态识别研究的一个热点,因此本文在多视角的背景下,重点研究了步态特征的提取以及视角归一化问题。本文完成的主要研究工作如下:首先,采用背景差分法完成目标的提取。针对目标图像中含有噪声及不连通等问题,对其采用形态学方法进行去噪,并进行连通区域分析来获得较完整的二值化步态图像。由于步态是一个周期性的运动,一个周期内的步态变化更能反映人体的运动特性,因此本文根据步态轮廓宽度变化来检测步态周期,并对步态图像进行标准化处理。其次,人体的步态图像序列不仅含有静态的步态信息,同时相邻步态间的变化也隐含了丰富的动态信息,而常用的步态能量图和步态帧差能量图只考虑了其中的静态信息和部分动态信息,因此本文将刻画不确定性的熵引入到步态帧差能量图中,提出采用步态帧差熵图刻画步态特征,再采用最近邻分类法完成分类识别。最后,针对多视角步态识别过程复杂、计算量大等问题,本文提出基于低秩优化的视角归一化方法进行步态识别。在步态特征图像的基础上,将任意视角下的步态特征图像采用秩优化的方法归一化到秩最小的视角,再采用最近邻分类法进行跨视角下的识别验证。实验结果表明,本文提出的步态帧差熵图的识别率要高于步态能量图和步态帧差能量图,而基于步态帧差熵图的视角归一化步态识别方法也在一定程度上提高了跨视角下的识别率。
[Abstract]:With the gradual deepening of intelligence, the security situation in public places is becoming increasingly serious. How to ensure public safety has become the focus of concern and attention, and effective identification technology is the key to ensure public safety. Biological features are widely used in the field of intelligent surveillance because of their security, stability and reliability. Compared with face, fingerprint and so on, gait has attracted a lot of researchers' attention because of its advantages of non-invasive, long-distance and difficult to hide. In recent years, gait recognition under multiple angles has been a hot topic in gait recognition. Therefore, this paper focuses on gait feature extraction and angle normalization under the background of multi-angle. The main work of this paper is as follows: firstly, the background difference method is used to extract the target. Aiming at the problems of noise and disconnection in the target image, the morphological method is used to de-noise and the connected region is analyzed to obtain a complete binary gait image. Gait is a periodic movement, and the gait changes in a period can reflect the movement characteristics of human body. Therefore, the gait period is detected according to the gait contour width change, and the gait image is standardized. Secondly, the gait image sequence of human body contains not only static gait information, but also rich dynamic information in the change of adjacent gait. However, the gait energy map and gait frame difference energy graph only consider the static information and some dynamic information, so the entropy which depicts uncertainty is introduced into the gait frame difference energy graph. Gait frame difference entropy graph is used to describe gait features, and the nearest neighbor classification method is used to complete classification and recognition. Finally, aiming at the complex process of multi-view gait recognition and the large amount of computation, this paper proposes a new method of gait recognition based on low-rank optimization. On the basis of gait feature image, the gait feature image with arbitrary angle of view is normalized to the angle of view with minimum rank by rank optimization method, and the nearest neighbor classification method is used to verify the recognition of gait feature image under cross-angle. The experimental results show that the recognition rate of the proposed gait difference entropy map is higher than that of gait energy map and gait frame difference energy map. The normalized gait recognition method based on gait frame difference entropy also improves the recognition rate of cross-angle gait to some extent.
【学位授予单位】：西安科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】