【作者】 袁国武；

【导师】 徐丹；

【作者基本信息】 云南大学 ， 通信与信息系统， 2012， 博士

【摘要】 智能视频监控就是利用计算机视觉和数字图像处理的方法,在不需要人为干预的情况下,对摄像机拍摄的图像序列进行自动分析,实现对动态场景中运动目标的检测、跟踪和识别,并在此基础上分析和判断目标的行为,给出对运动目标行为和动作的描述,自动发现一些可疑情况,实现系统对场景中的异常进行鉴别并自动报警,从而指导和规划人们的行动。目前,智能视频监控方面的研究和应用都面临着很多难题,国内外的许多学者投身于该领域的研究,取得了大量的成果。本文在这些成果的基础上,对智能视频监控系统中的关键步骤——运动目标检测与跟踪进行了研究。主要的工作概括如下：(1)针对运动目标检测和跟踪算法中的应用,改进了LBP (Local Binary Pattern)算子。LBP是一种强大的纹理描述算子,论文经过对LBP的二进制位串的01跳变情况进行统计,合并了出现概率较低的模式,使得LBP纹理的种类得到极大地削减,使后续使用LBP纹理的运动目标检测与跟踪算法在特征匹配时的速度得到提高。(2)提出了一种结合改进的LBP纹理和色度信息的运动目标检测算法。运用LBP纹理和色度这两种特征对阴影都不敏感的特性,论文用LBP纹理和色度信息对背景进行描述,然后把该背景描述应用到高斯混合模型中,提出了一种新的运动目标检测方法,该方法能较好的抵御阴影及背景光照变化的影响；(3)提出了一种结合光流法与三帧差分法的运动目标检测算法。由于计算光流的算法复杂,论文简化了光流的计算：选择图像中少数具有某些特征的像素点,只对这部分像素点计算光流信息,有效地减少了复杂度。但是,由于只选取了部分代表性像素,检测得到的运动目标区域不太完整,于是论文引入了三帧差分法作为简化光流法的补充,在不增加太多计算量的情况下,得到了相对完整的运动目标轮廓。(4)提出了联合LBP纹理和色度信息的Camshift跟踪算法。由于传统的Camshift算法是基于色彩直方图的目标跟踪,当目标和背景颜色相近或干扰目标和被跟踪目标颜色相近时,或者运动目标存在阴影时,跟踪的正确性受到很大的影响。论文采用LBP纹理和色度信息相结合描述运动物体,两种特征都对阴影不敏感,能克服阴影对跟踪结果的影响。另外,对于纹理缺乏区域,LBP纹理性能不好,但是色度往往能取得较好效果；而对运动目标和背景颜色相同的区域,而LBP纹理往往又能取得一定的效果。(5)提出了基于Kalman滤波器和Blob匹配法的目标跟踪算法。Blob匹配法利用目标的外形特征去匹配候选目标,该方法在运动物体数目较多时,每个Blob都要与其它的各物体相匹配,速度会变慢,且不适用于对非刚性物体的跟踪,同时在目标遮挡时也难于继续跟踪到目标。而Kalman滤波器可以根据目标先前的运动信息去预测下一帧中目标的位置,和Blob匹配法结合后,能在Blob搜索时只需在预测到的位置附近做Blob匹配,减少了计算量。论文得到了云南省应用基础研究计划项目“智能视频坚控中的运动目标检测与跟踪技术研究”(编号：2011FB019)的支持。更多还原

【Abstract】 Intelligent visual surveillance system can automatically analyze video sequences by the methods of computer vision and digital image processing without human intervention. The system can real-time detect, track and recognize moving objects in a scene. Furthermore, it can analyze and judge moving objects’ behavior, give a description of the behaviors and actions, automatically discover some suspicious behaviors, identify irregular actions in the scene and alarm automatically in order to guide our actions and decisions.At present, there are many problems in the research and application of intelligent visual surveillance. Both domestic and overseas scholars devote themselves to the research and get a large number of achievements. On the basis of these achievements, this thesis researches the technologies of moving objects detection and tracking, which are critical steps in the intelligent visual surveillance. The main work is summarized as follows:(1) The LBP (Local Binary Pattern) operator is improved. LBP is a powerful mean of texture description. After counting the times of0jumping to1and1to0in a LBP’s binary codes, the models, which have lower probability of occurrence, are combined. Therefore, the species of LBP texture are reduced, and the speed of feature matching is increased in moving objects detection and tracking algorithm based on the LBP texture.(2) A moving objects detection method based on a combination of improved local binary pattern texture and hue is proposed. Because both LBP texture and hue are not sensitive to shadows, they are used to describe a background. Then, the background is applied to the Gaussian mixture model, and the method can better resist shadows and the changes of background illumination.(3) A moving object detection algorithm based on a combination of optical flow and the three-frame difference is proposed. Because of the complexity of computing optical flow, the calculation of optical flow is simplified. A few pixels with certain characteristics are selected to compute optical flow information, which reduce the algorithm’s complexity. However, because of selecting parts of representative pixels only, target area detected by the algorithm is not complete, so three-frame difference method is introduced as a supplement in order to get a relative complete target area without increasing too much computing complexity.(4) A Camshift tracking algorithm based on LBP texture and hue is proposed. Traditional Camshift algorithm is based on the color histogram of moving objects. When an object and its corresponding background have similar color, an interferential object and the tracked object have similar color or moving objects have shadows, the tracking accuracy can be greatly affected. The LBP texture and hue are combined to describe moving objects in order to solve these troubles because two features are not sensitive to the shadow. In addition, in the region with similar color between moving objects and background, LBP texture can often achieve certain effects; and in the region of lacking texture, the LBP texture performance is not good, but color can often achieve better results.(5) A moving object tracking algorithm based on Kalman filter and blob matching is proposed. Blob matching method matches the candidate target using the moving objects’shape characteristics. When there are a large number of moving objects, the method has slower speed because each blob must match with other objects. Meanwhile, the method does not apply to non-rigid object tracking, and it is difficult to continue to track objects if the objects’occlusion is existed. Kalman filter can predict the target’s location in the next frame using the current frame’s motion information. Before blob matching, Kalman filter is used, so the matching area can be limited in the predicted area by Kalman filter.This thesis is supported by the application foundation project of Yunnan province "Research on moving objects detection and tracking methods in intelligent visual surveillance system"(No.2011FB019).更多还原