视频监控中的运动目标跟踪与识别研究

发布时间：2018-03-14 18:43

本文选题：智能视频监控系统　切入点：运动目标检测　出处：《电子科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：智能监控视频作为新一代的安防手段,使用计算机代替人工的工作,对视频序列进行检测、分析、跟踪和识别。随着“智慧城市”地推行,大量的数字视频信息,仅靠人工是无法完成信息分析的工作,可见对智能监控视频技术的研究是十分必要的。计算机视觉技术涉及到很多领域相关知识,如图像处理、模式识别、人工智能等,目前各个领域的研究已有众多成果,而监控视频所应用的复杂环境,并不能简单照搬各领域成果,其需求与科研需求并不是完全一样。目标跟踪与识别需要系统的有良好的实时性和稳定性,能在各种复杂场景中都可以快速准确的完成处理,所以鲁棒的、实时的算法是目标跟踪和识别算法研究的重点问题。本文根据监控视频的需求进行目标跟踪和识别技术的研究,主要的工作如下:分析对比各种运动目标检测的算法,提出了一种以背景差为辅助的三帧差法目标检测算法。该算法利用帧差法实时性好、适应能力强等特点,提出基于帧差法的分块背景提取法,以改善混合高斯模型背景提取计算复杂、实时性差的问题。针对帧差法在处理颜色相近的物体时易出现空洞,使用背景差法辅助,设置阈值对帧差法检测性能进行判断,当帧差法性能差时,使用背景差法。经过实验证明该方法相比传统算法在检测率和实时性都有提高。针对监控视频的需求,提出了一种LBP(Local Binary Pattern)纹理和色度相结合特征直方图,能很好的抑制监控视频中光照问题和遮挡问题。监控视频中物体大小会由远及近的变化,观测角度会有部分变化,使用Camshift算法,来解决这些问题对国内外识别技术进行分析,将HOG(Histogram of Oriented Gradient)特征应用到基于特征学习的目标物体识别算法,采集数据对弱分类器进行训练,级联分类器,对测试数据进行识别分类,分析实验结果。综上所述,本文基于监控视频的需求,提出了新的目标检测算法,完成了效果较好的目标跟踪和识别算法,可以应用于城市监控,智能交通等各种监控环境。
[Abstract]:Intelligent surveillance video as a new generation of security means, using computers instead of manual work to detect, analyze, track and identify video sequences. With the implementation of "smart city", a large number of digital video information, The research of intelligent surveillance video technology is very necessary. Computer vision technology involves many related knowledge, such as image processing, pattern recognition, artificial intelligence and so on. At present, there have been many achievements in various fields, but the complex environment used in surveillance video can not simply copy the achievements in each field. The requirement of target tracking and recognition is not exactly the same as that of scientific research. Target tracking and recognition need good real-time and stability of the system, and can be processed quickly and accurately in all kinds of complex scenes, so it is robust. Real-time algorithm is the key problem in the research of target tracking and recognition algorithm. According to the requirements of surveillance video, this paper studies the technology of target tracking and recognition. The main work is as follows: analysis and comparison of various moving target detection algorithms, In this paper, a method of target detection based on the background difference is proposed, which makes use of the advantages of the frame difference method, such as good real-time ability and strong adaptability, and proposes a block background extraction method based on the frame difference method. In order to improve the complexity of background extraction of mixed Gao Si model and the problem of poor real-time performance, the frame difference method is used to determine the detection performance of frame difference method, because it is easy to appear holes when dealing with objects of similar color, and the background difference method is used to assist, set a threshold to judge the detection performance of frame difference method. When the performance of frame difference method is poor, the background difference method is used. It is proved by experiment that the detection rate and real-time performance of this method are improved compared with the traditional algorithm. In order to meet the requirements of surveillance video, a LBP(Local Binary pattern histogram combining texture and chroma is proposed. It can restrain the illumination problem and occlusion problem in the surveillance video very well. The object size in the surveillance video will change from far to near, and the observation angle will change in part. The Camshift algorithm is used to solve these problems and analyze the recognition technology at home and abroad. The HOG(Histogram of Oriented gradient feature is applied to the target object recognition algorithm based on feature learning. The collected data is trained to weak classifier, cascaded classifier is used to identify and classify test data, and the experimental results are analyzed. Based on the requirement of surveillance video, a new target detection algorithm is proposed in this paper, which can be used in various monitoring environments such as urban monitoring, intelligent transportation and so on.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【相似文献】