无重叠视域下行人再识别算法的研究
发布时间：2018-08-20 12:18
【摘要】：近年来,随着人们自身公共安全意识的提高以及视频监控技术的发展,智能视频监控系统得到了大量的普及。行人再识别(Person Re-identification)是近几年智能视频分析领域兴起的一项新技术,是多摄像机联合智能视频监控系统中需要解决的关键问题之一,因而得到了广大计算机视觉领域及人工智能领域相关人员的关注。无重叠视域下的行人再识别是指在一个多摄像机联合的视频监控系统下,判断一个摄像头中出现的行人目标是否在另一个非重叠视域监控下的摄像头中出现过的一个过程,即识别出不同摄像机拍摄到的属于某一个行人的图像。但由于受摄像机角度、背景变化、光照条件、姿态变化、遮挡等多种外在复杂因素的影响,同一行人在不同视域下可能存在很大的差异性,从而使得行人再识别问题具有很大的挑战性。本文提出了一种基于度量学习的行人再识别算法PRML(Person Re-identification based on Metric Learning),主要通过特征学习生成一个测度矩阵来进行行人的再识别。本文首先通过一种图像增强算法对原始行人图像进行处理,从而减少因光照变化所带来的影响,然后根据人体目标外观形态特性对行人进行合理分割,并提取行人图像颜色特征(HSV、Lab)、纹理特征(SILTP、FHOG)以及颜色属性ColorNames特征并进行核函数学习,将原始线性特征空间投影到更加具有区分性的非线性特征空间并对特征进行PCA降维,之后考虑到不同类型特征对行人图像描述的差异性,分别学习得到三个独立的测度矩阵,并通过正则化方法来优化测度矩阵的过拟合问题,最终并加权融合多个测度矩阵综合得到行人图像对的相似性度量函数,从而实现行人相似性的度量。最后在VIPeR、iLIDS、CUHK01三个公共数据集上采用CMC(Cumulative Matching Characteristic Curve)曲线评测标准对提出的算法进行了实验效果验证、对比和分析。
[Abstract]:In recent years, with the improvement of public safety awareness and the development of video surveillance technology, intelligent video surveillance system has been widely used. Pedestrian rerecognition (Person Re-identification) is a new technology emerging in the field of intelligent video analysis in recent years. It is one of the key problems to be solved in multi-camera joint intelligent video surveillance system. As a result, the field of computer vision and artificial intelligence related to the field of attention. Pedestrian reidentification without overlap is a process in which a video surveillance system with multiple cameras is used to determine whether the pedestrian target in one camera has appeared in another camera. That is to identify the different cameras of the image of a pedestrian. However, due to the influence of many complicated external factors, such as camera angle, background change, illumination condition, attitude change, occlusion and so on, the same line of people may have great differences in different visual fields. Therefore, the problem of pedestrian recognition is very challenging. In this paper, a new pedestrian rerecognition algorithm based on metric learning (PRML (Person Re-identification based on Metric Learning),) is proposed, which is based on feature learning to generate a measure matrix for pedestrian rerecognition. In this paper, the original pedestrian image is processed by an image enhancement algorithm, so as to reduce the influence caused by the change of illumination, and then the pedestrian is segmented reasonably according to the appearance and morphological characteristics of the human object. The color feature (HSV), texture feature (SILTPFHOG) and color attribute (ColorNames) of pedestrian image are extracted, and the kernel function is studied. The original linear feature space is projected into a more discriminative nonlinear feature space and the feature dimension is reduced by PCA. Considering the difference of pedestrian image description between different types of features, three independent measure matrices are obtained, and the overfitting problem of measure matrix is optimized by regularization method. Finally, the similarity measurement function of pedestrian image pairs is obtained by combining multiple measure matrices weighted and weighted, and the pedestrian similarity measurement is realized. Finally, the experimental results of the proposed algorithm are verified, compared and analyzed by using CMC (Cumulative Matching Characteristic Curve) curve evaluation standard on three common data sets of VIPeR iLIDSU CUHK01.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6
本文编号：2193585
[Abstract]:In recent years, with the improvement of public safety awareness and the development of video surveillance technology, intelligent video surveillance system has been widely used. Pedestrian rerecognition (Person Re-identification) is a new technology emerging in the field of intelligent video analysis in recent years. It is one of the key problems to be solved in multi-camera joint intelligent video surveillance system. As a result, the field of computer vision and artificial intelligence related to the field of attention. Pedestrian reidentification without overlap is a process in which a video surveillance system with multiple cameras is used to determine whether the pedestrian target in one camera has appeared in another camera. That is to identify the different cameras of the image of a pedestrian. However, due to the influence of many complicated external factors, such as camera angle, background change, illumination condition, attitude change, occlusion and so on, the same line of people may have great differences in different visual fields. Therefore, the problem of pedestrian recognition is very challenging. In this paper, a new pedestrian rerecognition algorithm based on metric learning (PRML (Person Re-identification based on Metric Learning),) is proposed, which is based on feature learning to generate a measure matrix for pedestrian rerecognition. In this paper, the original pedestrian image is processed by an image enhancement algorithm, so as to reduce the influence caused by the change of illumination, and then the pedestrian is segmented reasonably according to the appearance and morphological characteristics of the human object. The color feature (HSV), texture feature (SILTPFHOG) and color attribute (ColorNames) of pedestrian image are extracted, and the kernel function is studied. The original linear feature space is projected into a more discriminative nonlinear feature space and the feature dimension is reduced by PCA. Considering the difference of pedestrian image description between different types of features, three independent measure matrices are obtained, and the overfitting problem of measure matrix is optimized by regularization method. Finally, the similarity measurement function of pedestrian image pairs is obtained by combining multiple measure matrices weighted and weighted, and the pedestrian similarity measurement is realized. Finally, the experimental results of the proposed algorithm are verified, compared and analyzed by using CMC (Cumulative Matching Characteristic Curve) curve evaluation standard on three common data sets of VIPeR iLIDSU CUHK01.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6
【参考文献】
相关期刊论文 前5条
1 齐美彬;檀胜顺;王运侠;刘皓;蒋建国;;基于多特征子空间与核学习的行人再识别[J];自动化学报;2016年02期
2 杜宇宁;艾海舟;;基于统计推断的行人再识别算法[J];电子与信息学报;2014年07期
3 范彩霞;朱虹;蔺广逢;罗磊;;多特征融合的人体目标再识别[J];中国图象图形学报;2013年06期
4 高勇;;大力推进视频监控建设与应用 不断提升公安工作能力和水平——全国公安机关视频监控建设与应用工作会议举行[J];中国安防;2012年08期
5 张志成;;基于监控视频的分析技术已成为执法调查的主要手段[J];中国安防;2012年08期
，本文编号：2193585
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2193585.html
