小区监控视频分析技术的研究与实现

发布时间：2018-03-24 07:00

本文选题：视频分析技术　切入点：视频动静态切分　出处：《华侨大学》2017年硕士论文

【摘要】：随着视频监控体系的日渐壮大,监控视频的内容也在呈指数级的增长,如何高效的管理、查看和检索这些视频内容成为一个急需解决的问题。小区监控视频存在图像质量差,光照变化大,设计要求经济实用等特点,本文主要针对小区监控视频分析中采用的前景检测、多目标跟踪和图像检索技术进行了研究和改进,并对这些技术在真实小区监控视频上进行实验对比,主要研究内容包括:(1)由于视频图像存在较多噪声影响后续处理,对常用的图像预处理技术进行了理论分析和实验对比,给出了最适合真实的小区视频数据的预处理算法。在视频动静态切分过程中,本文提出了一种基于行人运动规律的分块多阈值帧间差分算法,在真实的小区视频数据上的实验表明,该算法相对于普通帧间差分算法具有更快的切分速度和更低的漏检率。(2)由于小区监控的实际场景中存在大量的光照突变和光照不均匀现象,本文结合基于双背景模型的思想和一种自适应的光照补偿策略提出了一种能够应对光照变化的快速鲁棒前景检测算法。在Wallflower数据集和真实的监控视频数据上进行前景检测,实验结果表明该算法在发生光照突变时比混合高斯模型(GMM)、VIBE算法、自适应自组织网络算法(LBASOM)具有更高的检测精度,同时该算法相对于GMM、LBASOM和多层背景模型(MLBGS)具有更快的处理速度。(3)由于现有的大部分多目标跟踪算法计算复杂度过高,本文提出了一种结合meanshift算法和目标特征匹配的在线多目标跟踪算法,该算法无需训练模型,基本上不需要先验知识。在真实的小区监控视频数据上,相较于CMOT算法具有更低的跟丢率和更快的处理速度。(4)对颜色直方图、颜色和边缘方向特征描述符、图像哈希和SURF特征四种目标特征进行了理论分析,并对不同的目标特征所采用的相似性度量方法进行了改进,给出每种特征得到一个0-1之间的相似性度量值的计算方法,改进后的方法更加适用于图像检索。(5)将小区监控视频分析系统的分为动静态切分与信息提取层、多目标跟踪分析层和视频目标检索层。结合小区监控的实际需求分层实现了一个小区监控视频分析检索系统。
[Abstract]:With the development of video surveillance system, the content of video is growing exponentially. How to efficiently manage, view and retrieve these video content has become an urgent problem. The illumination change is large and the design requirement is economical and practical. In this paper, the foreground detection, multi-target tracking and image retrieval technology used in the video analysis of residential surveillance are studied and improved. The main research contents include: (1) because there is more noise in the video image to affect the follow-up processing, the common image preprocessing technology is analyzed theoretically and compared with the experiment. A preprocessing algorithm for real cell video data is presented in this paper. In the process of dynamic and static video segmentation, a multi-threshold inter-frame difference algorithm based on pedestrian motion is proposed. Experiments on real cell video data show that, The algorithm has faster segmentation speed and lower miss rate than the ordinary inter-frame differential algorithm. In this paper, a fast robust foreground detection algorithm based on dual background model and an adaptive illumination compensation strategy is proposed, which can deal with the change of illumination. It can detect the foreground on the Wallflower data set and the real surveillance video data. The experimental results show that the algorithm has higher detection accuracy than the hybrid Gao Si model GMMMM-VIBE algorithm and the adaptive ad hoc network algorithm (LBASOM). At the same time, the algorithm has a faster processing speed than GMM LBASOM and MLBGS) because of the high computational complexity of most of the existing multi-target tracking algorithms. In this paper, an online multi-target tracking algorithm combining meanshift algorithm with target feature matching is proposed. This algorithm does not require training model and requires no prior knowledge. The color histogram, color and edge direction descriptor, image hash and SURF feature are analyzed theoretically compared with CMOT algorithm. The similarity measurement method used in different target features is improved, and the method of calculating a similarity measure between 0-1 for each feature is given. The improved method is more suitable for image retrieval. 5) dividing the video analysis system into dynamic and static segmentation and information extraction layer. The multi-target tracking analysis layer and video target retrieval layer are used to realize a video analysis and retrieval system based on the actual requirement of cell monitoring.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】