复杂背景下动态目标的检测与跟踪

【作者】 王柱；

【导师】 李勃；

【作者基本信息】 昆明理工大学 ， 计算机软件与理论， 2007， 硕士

【摘要】 动态目标检测与跟踪是计算机图像和视频处理的基础,广泛地应用在工业、医学、军事、教育、商业、体育等领域中。本文针对常用智能监控系统中的动态目标检测与跟踪功能难以兼顾可靠性与实时性的缺点,着重讨论和研究了视频监控系统中运动目标检测和目标跟踪的关键技术,并研究了算法在普通个人电脑上的实现。在运动目标检测方面,本文研究和分析了大量的算法,对基于高斯统计模型的背景减法进行了改进,同时提出了一个将普通直方图建模方法与混合高斯模型建模方法相结合的复合检测框架。该框架可以对单模态的背景点建立直方图模型,对多模态的点建立混合高斯模型;另外,本文采用了彩色背景差分方式获取动态目标,并加入了阴影检测等功能使动态目标的获取更加完整,鲁棒性更强。目标跟踪方面,本文采用了一种简单的MLE(Maximum Likelihood Estimation,最大可能性评估)应用来分类目标,为多目标订立了跟踪优先级;在研究了目前常用的几种跟踪算法的基础上,改进了扩展卡尔曼跟踪算法,利用运动检测结果,通过运动区域的质心位置和面积比较来代替复杂的匹配过程,提高了跟踪的运算速度,同时简单有效的解决了复杂运动的情况。为了验证本文中模块算法的有效性,本文为各个模块设计了独立的实验,并采集了大量的视频数据进行对比,实验证明了算法的有效性。最后,本文在上述模块的基础上设计了一个完整的运动目标检测与跟踪系统的软件框架,实现了一个基于个人电脑的实时运动目标检测跟踪系统,并采用该系统进行了综合性能实验,系统实验获得了成功。更多还原

【Abstract】 Moving target detection and tracking are foundation of computer images and video processing work, which widely used in industrial, medical, military, education, business, sports and other areas. In order to solve the inadequacies that the moving target detection and tracking functions commonly used in intelligent monitoring system are difficult to balance reliability and real-time shortcomings, this dissertation discussed and studied the key technology of moving object detection and tracking on video surveillance systems, as well as the ordinary PC implementation of these algorithms.On the research of moving object detection, by studying and analyzing a great number of methods, in this dissertation, an improved background subtraction algorithm based on Gaussian model has been proposed, and a hybrid detection framework by combining ordinary Histogram model and Gaussian model differencing algorithm has been presented. This framework can establish Histogram model for background pixel which is single mode, and establish mixture Gaussian model for background pixel which is multiple mode. In addition, the paper used the background color difference method to get moving targets, and joined the shadow detection, and other functions so that the goal of moving targets dynamic acquisition more complete, robust stronger. This algorithm can achieve higher sensitivity and more robust detection result. On the research of object tracking, a simple version of MLE(Maximum Likelihood Estimation) is used to make the classification decision in this dissertation and built tracking priorities for multiple targets. By studying several related algorithm, this dissertation improves the extend Kalman filter to tracking targets which make use of the detection results and compare centroid position and area instead of complicated template matching, thus this method improve the tracking speed, as well as handling the cluttered scene with a simplicity and efficient approach.To validate the effects of these algorithms, this dissertation designed separate experiment for every module and collected a lot of video data. These experiments have proved the validity of these algorithms.Last, in this dissertation, a whole moving targets detection and tracking framework is designed, and a moving targets detection and tracking system on ordinary PC was implemented, which is executed successfully.更多还原