视频浓缩系统中的行人目标再辨识技术

发布时间：2018-07-28 18:53

【摘要】：随着人们安防意识的提高以及数字视频技术的提升,视频监控网络已成为维护社会治安的一个重要组成部分。视频浓缩系统能对长时间、大范围的监控视频进行浓缩摘要,提高信息挖掘的效率。在视频监控网络中常常面临对指定行人目标进行识别分析的需求,因此在视频浓缩系统中实现行人再辨识有着重要的现实意义。本文结合视频浓缩系统相关技术,对现有行人再辨识技术做出了一些改进,主要工作如下:第一,提出了一种无监督的实时行人再辨识方法。为应对海量监控视频行人目标图像,本文提出了一个两级搜索的框架:在一级搜索建库时,针对每个行人目标图像提取快速稳健的颜色、纹理等局部显著特征并进行降维处理。在检索时,首先,根据给定的一张查询行人目标图片,在数据库中进行快速线性相似性度量,筛选出容量较小的结果候选集;然后,二级搜索时再在候选集中对每张行人目标图像提取相对耗时的局部特征进行二次相似度匹配,重新排序得到最后的再辨识结果。为加快匹配速度,本文设计了一种提取速度较高、区分性较好的局部描述子并进行VLAD编码(MLD-VLAD)。通过相应对比实验表明,在行人再辨识技术中,与颜色空间相融合的MLD-VLAD取得比SIFT局部描述子更好的区分性能。第二,提出了一种有监督的实时行人再辨识方法。在无监督实时行人再辨识的基础上引入了大间隔近邻(LMNN)算法,并做出了相应改进。本文提出了基于皮尔逊相关距离的大间隔近邻算法(p-LMNN),旨在训练一个线性转换矩阵L,利用L对原始特征以降维的方式进行投影,在新的特征空间内进行距离度量。通过实验表明,p-LMNN要取得比LMNN更好的再辨识性能表现。第三,在实时行人再辨识的基础上,扩展出了实时汽车再辨识。由于汽车本身特征较为刚性,即形状、姿态变化较小,无监督的再辨识模式已经取得较好的表现。实时汽车再辨识模式有较好的推广泛化能力,更换应用场景无需人工采集样本进行标注,简单易用。
[Abstract]:With the improvement of people's security awareness and the improvement of digital video technology, video surveillance network has become an important part of maintaining public order. Video concentration system can concentrate the long-time and large-scale surveillance video and improve the efficiency of information mining. In video surveillance networks, it is often faced with the need to identify and analyze designated pedestrian targets, so it is of great practical significance to realize pedestrian re-identification in video concentration systems. In this paper, some improvements have been made to the existing pedestrian reidentification technology in combination with the video concentration system. The main work is as follows: first, an unsupervised real-time pedestrian reidentification method is proposed. In order to deal with the massive surveillance video pedestrian target images, this paper proposes a two-level search framework. In the first level searching and building database, the fast and robust local salient features such as color and texture are extracted and dimensionally reduced for each pedestrian target image. In the retrieval, first, according to a given query pedestrian target image, a fast linear similarity measure is carried out in the database, and the candidate set with small capacity is screened out. In the second level search, the local features of each pedestrian target image are extracted by quadratic similarity matching in the candidate set, and the final recognition result is obtained by reordering. In order to speed up the matching speed, this paper designs a local descriptor with high extraction speed and good discriminant and carries out VLAD coding (MLD-VLAD). The corresponding experiments show that the MLD-VLAD fusion with color space is better than the SIFT local descriptor in pedestrian reidentification. Secondly, a supervised real-time pedestrian re-identification method is proposed. Based on the unsupervised real-time pedestrian reidentification, a large interval nearest neighbor (LMNN) algorithm is introduced and improved accordingly. In this paper, a large interval nearest neighbor algorithm (p-LMNN) based on Pearson correlation distance is proposed, which aims at training a linear transformation matrix L, projecting the original feature in a dimensionality reduction way using L, and measuring the distance in a new feature space. The experimental results show that p-LMNN is better than LMNN in re-identification performance. Thirdly, on the basis of real-time pedestrian recognition, real-time vehicle re-identification is extended. Due to the relatively rigid characteristics of the vehicle, that is, the shape and attitude change is small, the unsupervised re-identification model has achieved better performance. The real-time vehicle reidentification model has better generalization ability, and it is simple and easy to use for changing the application scene without manually collecting samples for labeling.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】