基于视频的商场人群行为识别研究

发布时间：2018-05-26 01:50

本文选题：商场 + 人群行为　；参考：《天津财经大学》2016年硕士论文

【摘要】：随着当前社会经济的高速发展,人们的生活质量不断提高,大量的集购物、吃饭、玩乐等功能为一体的大型商场场所被建造以供给人们用来平日的休闲娱乐活动,满足人们的各种需求。视频监控被广泛地用于商场内监控人们的活动,保证商场的正常运营。但视频监控往往只被用于监控商场内人群是否存在异常,没有充分利用其价值。目前,基于视频的人群行为识别分析已是国内外广泛重视研究的热点。而对商场监控视频中人群一般正常活动时的行为进行识别研究,不仅具有重要的公共安全研究价值,而且对于如今受到电子商务冲击的传统商场零售业而言也有一定商业研究价值。根据目前研究发现,人群的运动往往不是直接由个体运动组成的,而是由大大小小的群组的行为组成的。因此,基于上述背景内容,对商场内人群通过分组来分析其正常行为,设计了一种基于视频的商场人群行为识别方法,对商场中普遍常见的人群自由前行、人群排队以及人群坐着休息等候三种正常行为进行识别。首先,对视频中的人群前景进行检测,提出一种结合YUV空间码本模型和基于HOG特征行人检测的改进前景人群目标检测算法。在使用YUV空间码本模型检测得到前景目标的基础上,结合基于HOG特征的行人检测结果,进一步准确地得到重点研究的前景人群目标。其次,在检测到的人群前景目标基础上提取人群特征点,采用金字塔LK光流法对人群特征点进行跟踪并计算光流信息,减少了直接使用光流法的计算量。同时,再使用基于HOG特征检测行人方法,检测到的行人区域结果将用于对人群分组。再次,将跟踪到的人群特征点进行分组,提出一种改进的k-means二次聚类算法。第一次先使用k-means聚类算法对检测出的行人区域中心点进行聚类,得到人群总体粗分组情况。在此基础上,将第一次聚类得到的人群粗分组的各中心,作为再次使用k-means聚类算法初始时的聚类中心,对所有人群特征点进行细划分组。并结合基于HOG特征检测行人的结果对分组结果加以调整,得到最终的人群特征点分组。最后,对每个群组的人群特征点提取运动特征和密度特征,采用支持向量机SVM的分类识别算法,在自行采集的和网上收集的视频数据集上进行训练和测试识别商场内三种人群正常行为,取得了较好的实验结果,实现了解整体人群的运动情况的目的。
[Abstract]:With the rapid development of social economy, people's quality of life is improving constantly. A large number of shopping, eating, playing and other functions are built to provide people with leisure and entertainment activities. Meet people's needs. Video surveillance is widely used in shopping malls to monitor people's activities to ensure the normal operation of shopping malls. However, video surveillance is often only used to monitor whether there are anomalies in shopping malls, and the value of video surveillance is not fully utilized. At present, video-based crowd behavior recognition and analysis has been the focus of extensive attention at home and abroad. However, it is not only of great value to study the behavior of the general normal activities of people in the surveillance video of shopping malls, but also to study the public safety. And it also has certain commercial research value to the traditional retail business which is impacted by e-commerce. According to current studies, the movement of people is usually composed not directly of individual motion, but by the behavior of large and small groups. Therefore, based on the above background, this paper analyzes the normal behavior of the people in the shopping mall by grouping, and designs a video-based method to identify the behavior of the crowd in the shopping mall. The crowd queue and the crowd sit down and wait for three normal behaviors to be identified. Firstly, this paper presents an improved foreground crowd detection algorithm based on YUV spatial codebook model and pedestrian detection based on HOG features. On the basis of using YUV spatial codebook model to detect the foreground target, combined with the pedestrian detection result based on the HOG feature, the foreground crowd target of the key research is obtained accurately. Secondly, based on the detected population foreground target, the crowd feature points are extracted, and the pyramidal LK optical flow method is used to track the crowd feature points and calculate the optical flow information, which reduces the calculation amount of the direct use of the optical flow method. At the same time, the pedestrian detection method based on HOG features is used, and the detected pedestrian area results will be used to group the population. Thirdly, an improved k-means quadratic clustering algorithm is proposed by grouping the tracked population feature points. For the first time, the k-means clustering algorithm is used to cluster the detected pedestrian center, and the total coarse grouping of the population is obtained. On this basis, the centers of the first clustering are used as the initial clustering centers of the k-means clustering algorithm to divide all the population feature points into groups. Combined with the result of pedestrian detection based on HOG feature, the result of grouping was adjusted to get the final grouping of crowd feature points. Finally, the motion feature and density feature are extracted from the crowd feature points of each group, and the classification and recognition algorithm of support vector machine (SVM) is used. Three kinds of normal behaviors of people in shopping malls are trained and tested on the video data set collected by ourselves and online. Good experimental results are obtained and the purpose of understanding the movement of the whole population is achieved.
【学位授予单位】：天津财经大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TN948.6

【相似文献】