监控视频下行人再识别技术研究

发布时间：2018-04-08 17:56

本文选题：行人再识别　切入点：超像素　出处：《电子科技大学》2016年硕士论文

【摘要】：随着视频监控的逐渐普及,监控视频系统在社会公共安全和维护稳定中发挥着越来越重要的作用,对于大部分的监控系统,视频中的人及人的行为都是关注和研究的重点对象,在视频侦查的应用中,对监控视频画面中的人进行检测、识别、跟踪和分析其特定行为,利用监控系统检测到的信息做进一步的分析,而其中行人再识别技术就是监控视频中的核心关键问题,它指的是判断不同摄像头下出现的行人是否属于同一行人。随着计算机视觉的发展和社会应用的不断需求,行人再识别这一问题正逐渐发展成热点,基于行人在监控摄像头下的特点,本文提出了三种基于行人外貌特征的再识别算法,主要的工作如下:首先,本文提出了一种基于超像素分割的行人再识别算法。先用图像填充的方法对检测到的前景行人图像进行填充,再将行人分割成超像素块,对每个超像素块提取其颜色特征,并根据事先训练好的行人颜色特征的聚类中心,计算与每个颜色特征最靠近的聚类中心,统计每个超像素块的聚类中心直方图作为该超像素块的特征,为了匹配不同的行人,本文提出了一种超像素块投影匹配的算法,重点解决了行人特征之间的匹配问题。然后,本文建立了单个监控视频下、不同时间段、不同远近的行人再识别系统框架,提出了基于行人部件分割的再识别算法,本文训练了行人不同部件的检测器,分别提取每个部件的颜色和纹理特征,将多个颜色空间级联起来作为颜色特征,融合局部二值模式(LBP)纹理特征,将行人之间的匹配问题转换成行人各部件之间的匹配问题。最后,本文建立了不同摄像头下、同一时间段、不同角度的行人再识别系统框架,提出了基于不同行人区域特征匹配的方法。本文将行人不同区域划分成大小相同的小块,对每个小块提取颜色均值,再对该行人与目标行人对应区域的所有颜色值进行聚类,提取聚类后的颜色直方图并且结合尺度不变特征变换(SIFT)描述子,计算对应行人区域之间距离进行行人匹配。
[Abstract]:With the popularity of video surveillance, video surveillance system plays an increasingly important role in social public security and maintenance and stability. For most surveillance systems, people and their behavior are the focus of attention and research.In the application of video detection, the detection, identification, tracking and analysis of the specific behavior of the people in the surveillance video picture are carried out, and the information detected by the monitoring system is used for further analysis.The technology of pedestrian rerecognition is the key problem in surveillance video, which means to judge whether the pedestrian under different cameras belongs to the same pedestrian.With the development of computer vision and the continuous demand of social application, the problem of pedestrian rerecognition is gradually developing into a hot spot. Based on the characteristics of pedestrians under the surveillance camera, this paper proposes three rerecognition algorithms based on the features of pedestrian appearance.The main work is as follows: firstly, a pedestrian rerecognition algorithm based on hyperpixel segmentation is proposed.First, the foreground pedestrian image is filled with the image filling method, then the pedestrian is divided into super-pixel blocks, and the color features of each super-pixel block are extracted, and the clustering center of the pre-trained pedestrian color feature is used.The clustering center closest to each color feature is calculated, and the cluster center histogram of each super-pixel block is counted as the feature of the super-pixel block. In order to match different pedestrians, a super-pixel block projection matching algorithm is proposed in this paper.The matching problem between pedestrian features is solved emphatically.Then, this paper establishes the framework of pedestrian rerecognition system based on single surveillance video, different time period, different distance and near, proposes a rerecognition algorithm based on pedestrian component segmentation, and trains the detector of different pedestrian parts.The color and texture features of each component are extracted, and several color spaces are concatenated as color features to fuse the local binary pattern (LBP) texture features, and the matching problem between pedestrians is transformed into the matching problem among pedestrian components.Finally, the framework of pedestrian rerecognition system with different cameras, the same time period and different angles is established, and a method based on different pedestrian area feature matching is proposed.In this paper, different areas of pedestrians are divided into small blocks of the same size, and the average color of each block is extracted, and then all the color values of the corresponding area between the pedestrian and the target pedestrian are clustered.The color histogram after clustering is extracted and the SIFT descriptor is used to calculate the distance between the corresponding pedestrian areas for pedestrian matching.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TN948.6

【相似文献】