【作者】 郝彦斌；

【导师】 胡家升；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2008， 硕士

【摘要】 本文重点研究了目前运动目标检测与跟踪领域的一些常用方法,以数字图像处理技术和数字信号处理理论为基础,利用DirectShow构建了一个数字图像采集系统,根据实际应用情况编制了运动目标检测与跟踪实验软件,为运动检测与跟踪提供了算法实验平台和优化了硬件设计方案。在运动目标检测方面,主要研究了帧间差分、背景差分和光流场的方法。本文根据所处理视频图像相邻帧具有相关性的特点,采用了帧间差分目标检测方法。在差分图像之前,先对图像进行滤波和平滑等降噪处理,再进行相邻帧差分以得到运动信息。在背景差分方法中设计了统计平均法及其改进型、中值滤波法、单高斯和多高斯混合模型法提取背景,针对不同复杂程度的背景使用不同的方法可以获得不同的效果。在介绍光流基本原理的基础上,提出了适合计算机快速运算的光流方程迭代解法。对检测目标灰度图像采用了直方图、最大熵、最大类间方差及自适应法多种阈值分割方法,将灰度图像转化成黑白二值图像。对目标粗糙的边缘轮廓和背景进行了数学形态学滤波,较好的除去背景噪声、提取目标轮廓,采用了投影法确定目标的位置和大小。提出了一种操作简便、运算速度快、检测精度高、易于移植到硬件实现的算法流程。在运动目标跟踪方面,本文详细研究了相关匹配的ABS、SSDA、规一化和金字塔分层匹配算法,针对多点相关跟踪算法运算量大、实时性差的缺点,构建了一种改进的自适应阈值序列的SSDA模型。对模板在图像中按螺旋方式从中心向外遍历,并且只将对应二值化模板图像中像素值为1的像素参与匹配计算,使得匹配运算点完全集中在目标像素上,减少了计算量、提高了匹配精度和速度,同时采用模板尺寸修正及动态模板更新的方法,保证了跟踪的准确性。还研究了粒子滤波、卡尔曼预测确定运动目标下一预定位置,使用彩色直方图特征匹配进行确认,通过实验进行了分析比较。提出差分检测与主动轮廓跟踪相结合的方法,减少了计算迭代次数、提高了计算速度。最后,对所研究算法进行了合理规化、优化建模,使用C++语言结合DirectShow组件,在Window XP环境的Visual Studio 2005平台上开发了基于视频序列图像运动目标检测与跟踪系统。通过该系统对研究场景进行实验,可以在算法的可靠性和实时性上得到量化比较结果,为设计应用系统提供了捷径。更多还原

【Abstract】 The thesis research some techniques which is common in the area of moving object detection and tracking recently.By using DirectShow.a digital image sampling system has been built on the basis of digital image and digital signal processing theory.A software using for detection and tracking objects experimental is developped according practical application,that provide a playform for testing algorithms and optimize hardware designing solution for moving detection and tracking.On moving object detection,main concern is consecutive frames difference、background difference and optical flow.Image is enhanced using fliter and smoothing forreducing noise before image difference.On background difference,the methods of statistical average、median filter、single Gaussian model and mixture Gaussian model arc designed to extracting the background.For various complexity background using different methods to getting different results.On describing basic principle of optical flow,provided an iterative algorithm for solution of optical flow equation,which is fitting for computer highspeed caculating.For gray image of moving object detection and transforming it to binary image,involved many threshold segmentation method such as histogram、maximum entropy、maximum between-cluster variance and adaptivemethod.Mathematical morphological filter is done to rough edges of object and back-ground efficiently reduced the background noise and extracted object contour.Using projection method to determine the location and size of object.A algorithm flow whichis simple and convenient for operating high-speed high quality of detecting precision and easily transplant for hardware implementation is presented.On the motion tracking,correlation matching algorithm such as ABS,SSDA, nor-malization and pyramid delaminating matching are deeply studied in the thesis.Due to large computing of multiple point correlation tracking algorithm and bad realtime capability, an improved SSDA model on adaptive threshold sequence is constructed.Template moves in a spiral mode from center of the image to the outward,pixel in the corresponding binary template whose value is 1 is calculated and make matching operational points concentrate on object pixels,thus it reduces caculation and improvesmatching precision and speed.Also it guarantees the accuracy by using the method of dimension correction on template and dynamic template updating.In addition,particlefilter and kalman predietion are studied to determine the expected position about moving object,color histogram feature matehing is introduced to make confirmation,then analysis and comparison are obained through experiment.A method on differential detection combined with active contour tracking is presented,which reduces iteration times and improves calculation speed.Finally,it does reasonable planning and optimization modeling to the algorithm.Byusing c++ language combined with DireelShow component,a system based on video sequence image moving target detection and tracking is developed on Visual Studio 2005 platform in Window XP enviroment.By experiment of the research,we can getquantization of the algorithm in respect of relability and realtime,simultaneous,a shortcut is provided to design application system.更多还原