基于监控视频的运动目标检测算法

发布时间：2018-04-06 05:10

本文选题：运动目标检测　切入点：高斯模型　出处：《吉林大学》2017年硕士论文

【摘要】：近年来,随着计算机技术的蓬勃发展,互联网技术也随之走进了千家万户,人们的日常生活越来越离不开各种智能设备,智能视频监控系统也应运而生。用计算机视觉代替人眼去识别视频中的目标,不仅能大大的提高工作效率,同时还能解决很多人眼解决不了的问题,在这样的前提下,视频中的运动目标检测算法变得越来越重要。一个好的运动目标检测算法,能够实时、准确的将运动目标从视频当中提取出来,不仅应用在人们日常生活的小区、车站等人群密集的地方,同时在交通、军事、医疗方面都有着广泛的应用,而且在将视频中的运动目标提取出来之后,我们还可以对其进行跟踪、行为分析等进一步的操作。因此,迅速又完整的提取出运动目标是视频挖掘分析的基础。本文首先介绍了光流法、帧间差分法和背景差分法这三种传统运动目标检测方法的基本原理,并分析了它们各自的优缺点,还介绍了平均值法、中值法、核密度估计法和单高斯模型这四种背景建模的方法,旨在通过对以上几种运动目标检测过程的分析,对不足之处做出改进,然后结合每种算法的优点,总结出一套适用度更加广泛、目标检测结果更加精准的运动目标检测算法。本文所阐述的算法主要是对帧间差分法做出改进,提出了改进之后的五帧差分法,使得运动目标的检测结果更加精细,同时还解决了边缘不够连续的问题,从而保证了信息的完整性;此外,还针对混合高斯模型中,计算量过于庞大,以及?取值只能固定不变的问题做出了改进,通过将视频图像分块运算,大大降低了匹配算法的运算量,同时还提出了一种新的方案,使?能够根据当前视频流过的帧数自适应取值,能够在视频刚开始之初,就迅速完成背景建模,之后为了更好的抑制噪声,再逐渐减小?的取值。由于帧间差分法对光照变化不敏感,但是对于连续迅速变化的动态背景,其运动目标检测结果就会大打折扣;而混合高斯模型恰恰适用于变换复杂的动态背景,却对光照的强烈变化异常敏感。因此,本算法融合了二者的优点,互相取长补短,将改进后的五帧差分法和改进后的混合高斯模型结合起来,共同完成监控视频中的运动目标检测。
[Abstract]:In recent years, with the rapid development of computer technology, Internet technology has also entered thousands of households, people's daily life is more and more inseparable from a variety of intelligent equipment, intelligent video surveillance system also came into being.Using computer vision instead of human eyes to identify video targets can not only greatly improve work efficiency, but also solve problems that cannot be solved by many people's eyes.Moving target detection algorithms in video are becoming more and more important.A good moving target detection algorithm can extract moving target from video in real time and accurately. It is not only used in the crowded places of people's daily life, such as community, station and so on, but also in traffic, military, etc.Medical applications are widely used, and after extracting moving targets from video, we can track them, analyze their behavior and so on.Therefore, the fast and complete extraction of moving targets is the basis of video mining and analysis.In this paper, the basic principles of three traditional moving target detection methods, optical flow method, inter-frame difference method and background difference method, are introduced, and their respective advantages and disadvantages are analyzed.The kernel density estimation method and single Gao Si model are four background modeling methods, which aim to improve the shortcomings of each algorithm by analyzing the above process of moving target detection, and then combining the advantages of each algorithm.A set of moving target detection algorithms with wider applicability and more accurate target detection results are summarized.The algorithm described in this paper is mainly to improve the method of inter-frame difference, and the improved five-frame difference method is put forward, which makes the detection result of moving target more precise, and solves the problem that the edge is not continuous enough.In order to ensure the integrity of the information; in addition, the mixed Gao Si model, the calculation is too large, and?The problem that the value can only be fixed is improved. By dividing the video image into blocks, the computational complexity of the matching algorithm is greatly reduced. At the same time, a new scheme is proposed.Can choose the value according to the frame number of the current video flow, can finish the background modeling quickly at the beginning of the video, and then reduce gradually in order to suppress the noise better?The value ofBecause the inter-frame differential method is not sensitive to the illumination change, but the moving target detection results will be greatly reduced for the continuous and rapidly changing dynamic background, and the mixed Gao Si model is suitable for the transformation of complex dynamic background.But it is very sensitive to intense changes in light.Therefore, this algorithm combines the advantages of the two methods and makes up for each other's weaknesses. The improved five-frame differential method is combined with the improved mixed Gao Si model to detect moving targets in surveillance video.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】