基于视频的人流量统计技术研究

发布时间：2018-04-05 06:53

本文选题：人流量统计　切入点：目标检测与跟踪　出处：《华中科技大学》2014年硕士论文

【摘要】：视频监控系统兴起于二十世纪七十年代，但发展到现在已经不只是单纯的场景记录功能，，随着计算机视觉、模式识别、机器学习等相关技术的发展和成熟，许多以往在视频监控系统中需要人工分析操作完成的任务已经能够通过这些技术自动化完成，因此更准确地说现在的视频监控系统是智能视频监控系统。基于视频的人流量统计技术正是智能视频监控系统中的一个重要应用。人流量统计数据是商场、地铁、机场等公共场所进行管理和决策的重要依据，因此人流量统计技术有着非常广泛的应用前景。基于视频的人流量统计技术通常包含目标检测、目标跟踪、目标计数等相关技术，本文提出了一种在已知视频场景参数下的基于HOG特征检测与跟踪的人流量统计算法，该算法采用检测场景中行人的人头部分来作为计数依据，避免了处理目标遮挡问题；使用HOG特征来描述目标表观特征，使得算法具有抗光照变化能力和较好的鲁棒性；采用检测与跟踪相融合的目标跟踪算法，高效地处理了场景中的多目标跟踪问题；计数实现了跨线计数和区域计数两种方式，满足了实际应用需求。针对人流量统计技术实际应用场景中的摄像头多角度问题，提出了一种基于多分类器解决方案，在不采用复杂分类器的情况下提升了目标检测的精度，并且给出了多分类器自选择算法，自动化的选择最佳分类器。视频中的行人统计目标大小是未知的，并且存在多种尺度，我们在文中提出了一种目标检测尺度在线学习的算法，能够自动学习出场景中各区域的最佳目标检测尺度，从而不再需要使用多尺度扫描检测方法，极大的加速了人流量统计算法的时间效率。
[Abstract]:Video surveillance system rose in 1970s, but now it is not just scene recording function. With the development and maturity of computer vision, pattern recognition, machine learning and other related technologies,Many tasks which need manual analysis and operation in video surveillance system can be completed automatically by these technologies. Therefore, the current video surveillance system is an intelligent video surveillance system.The human flow statistics technology based on video is an important application in intelligent video surveillance system.The statistical data of population flow is an important basis for management and decision making in public places such as shopping malls, subway, airport and so on. Therefore, the statistical technology of population flow has a very broad application prospect.The techniques of human flow statistics based on video usually include target detection, target tracking, target counting and so on. In this paper, we propose a new human flow statistics algorithm based on HOG feature detection and tracking under known video scene parameters.The algorithm uses the human head of the pedestrian in the scene as the counting basis to avoid the target occlusion problem, and uses the HOG feature to describe the target surface feature, which makes the algorithm has the ability to resist the change of illumination and has better robustness.A target tracking algorithm which combines detection and tracking is used to efficiently deal with the multi-target tracking problem in the scene, and cross-line counting and region counting are realized to meet the practical application requirements.Aiming at the problem of multi-angle camera in the practical application of human flow statistics technology, a multi-classifier based solution is proposed, which improves the accuracy of target detection without using complex classifier.The self-selection algorithm of multi-classifier and the automatic selection of optimal classifier are also given.The pedestrian size in video is unknown and there are many kinds of scales. In this paper, we propose an online learning algorithm of target detection scale, which can automatically learn the best target detection scale in each region of the scene.Therefore, it is no longer necessary to use multi-scale scan detection method, which greatly accelerates the time efficiency of the algorithm.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.6

【参考文献】