复杂环境下运动目标的检测与跟踪

【作者】 张绍星；

【导师】 任明武；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2008， 硕士

【摘要】 运动目标的检测和跟踪是计算机视觉、图像处理、模式识别等多领域共同研究的热门课题,近些年来已经取得的很多成果,被广泛地应用到军事导航、监控监视、医学诊断、视频检索等众多领域。在军事侦查等条件下环境复杂多变,目标纷繁多样,出没突然缺乏规律可循,所以侦查监视系统要以稳健的目标检测与跟踪算法为核心。本文重点探讨了复杂背景下运动目标检测和跟踪的方法。主要工作和贡献如下:1、研究了静态背景下目标检测方法和背景模型生成算法。分析了一种基于LBP纹理描述的背景建模方法,应用改进的LBP直方图对背景建模,收到较好的检测效果;通过试验分析了传统目标检测算法的特点,并对部分算法做出了相应改进。最后依据已经分析的目标检测算法的优缺点,采用一种基于对称差分和背景差分的综合运动检测算法,将对称差和多高斯背景以及单高斯背景检测结果进行了有效结合,以提高目标检测的完整性,克服了运动目标短暂滞留的漏检问题。2、研究了一种基于卡尔曼滤波的跟踪方法。分析了目标匹配算法、卡尔曼滤波和均值偏移算法,特别是卡尔曼滤波和均值偏移算法的理论基础及其在目标跟踪过程中的应用,并做出了对应跟踪方法的具体实现。3、研究了抗遮挡的目标跟踪方法。改进了一种基于运动预测的均值偏移跟踪算法,根据跟踪过程中运动目标受到遮挡与否,采用不同处理策略将卡尔曼滤波和均值偏移两种算法协同完成跟踪任务,该方法引入颜色和LBP纹理联合直方图的目标建模方法,代替原算法纯粹的颜色直方图建模方法,并引进目标的时空运动位置信息,从而提高了遮挡条件下目标跟踪的鲁棒性。最后研究了一种基于均值偏移的粒子滤波算法,采用颜色和LBP纹理联合直方图对目标建模,建立具有一定自适应能力的动态系统模型,通过均值偏移对粒子的一步迭代,使粒子相对聚类到高概率位置,达到了减少粒子数目提高算法实时性的目的。更多还原

【Abstract】 The detection and tracking of moving object is a very popular subject in the area of Compute Vision、Image Processing、Pattern Recognition and etc, Lots of achievements had been made in recent years. It has been widely applied in military missile guidance、monitor and surveillances medical diagnosing、video retrieval and other fields. And that under the condition of battleground, the environment is very complex and varied , there are kinds of targets that appear and disappear with few rules. So the kernel of monitor and surveillance system is the robust algorithms of moving target detection and tracking. The paper mainly probe into methods of detection and tracking in complex environment. The major works and innovation of this paper include:1、Studying moving target detection and Background Model built under static background. We analyze the character of each algorithm in target detection by experiment, some algorithms have been improved. Paper improved a novel algorithm of background model based on LBP texture, and getting a well detecting result. Aimed at the advantage and disadvantage of each method on target detection analyzed before, we adopt an integrated detecting method combined the Symmetrical difference and Background Subtraction in the moving region. Using Symmetrical difference、Multi-gauss Background and Single-Gauss Background combined availably, improved the integrality of detection, conquered the miscarriage of justice of target’s whistle stop.2、Studing a tracking method based kalman filtering. Analyzing target matching、kalman filtering and mean shift algorithms, Particularly the wide application in moving object tracking of kalman filtering and mean shift. Then realized the correlative tracking methods of them.3、Studying the moving target tracking methods resisting the occlusion. An improved mean shift algorithm based on moving prediction has been proposed, which import target modeled with color and LBP texture united histogram, using real-time object’s position information to make target tracking more robust. Lastly paper propose an improved particle filtering algorithm based on mean shift, which adopt color and LBP texture united histogram-based target model, build a self- adaptively dynamic system model, use mean shift iterative to every particle to high probability position, so that reducing the number of particles and improving the real-time quality.更多还原