【作者】 焦波；

【导师】 李国辉；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2009， 博士

【摘要】 随着固定摄像机视频监控系统的广泛应用,面对海量监控视频数据,人们不仅需要有效管理,还需要能够24×7全天候实时自动从中提取出感兴趣的信息和知识,实现监控视频的智能化。运动目标的检测与跟踪是智能视频监控系统中最基础的两项核心技术,它们是视频监控技术智能化和实时应用的关键。传统运动目标检测与跟踪方法通常针对特定场景设计,难以应对各种复杂环境的变化,并且在准确性和实时性之间难以达到较好的折中。针对上述需求,本文以单目固定机位的摄像机(固定摄像机)输出的视频图像序列为研究对象,以视频运动目标的检测与跟踪及其相关的关键技术为研究内容,以期实现实时精确的运动目标检测和跟踪方法。其主要工作如下:(1)根据监控视频背景模型的复杂程度,将固定摄像机视频监控系统的应用场景划分为简单场景和复杂场景两类。针对简单场景提出一种基于区域划分的运动目标检测方法,提高了传统方法的检测准确性和光照低敏感性。针对复杂场景提出一种用于运动目标检测的快速收敛混合高斯模型,在确保准确性的前提下,提高了传统方法的收敛速度和时间效率。(2)运动目标检测获得区分背景区域与前景目标区域的二值图像,为了实现前景目标区域中不同目标的分割及噪声消除,提出一种基于链路的视频序列中二值图像的快速聚类方法,并提出一种目标内部空洞的快速填充方法,弥补了传统方法时间效率低、时间效率不稳定及聚类和空洞填充效果差的缺陷。(3)监控视频监控对象通常是人或车辆,采用人的3D竖直椭圆体模型可以较好的解决人的部分遮挡及阴影问题,但由于车辆模型的复杂性,采用车辆模型解决部分遮挡及阴影问题将耗费大量时间,针对这一问题,提出一种基于形态学的部分遮挡车辆分割及阴影消除方法,在确保准确性的前提下,显著提高了传统方法的时间效率,能满足现有硬件条件下的实时性。(4)提出一种基于自适应粒子滤波的目标跟踪方法,解决现有粒子滤波对固定运动模型依赖性强的缺陷,提高了传统方法的准确性。通过快速有效运动目标轮廓、几何形态和颜色特征的提取,提出一种预测与特征匹配相结合的目标跟踪方法,提高粒子滤波跟踪的准确率。本文在准确性和实时性两个标准下,验证了各种方法的有效性。这些研究将为固定摄像机监控视频中的运动目标检测与跟踪技术做出有益探索。更多还原

【Abstract】 With the widely application of static camera video surveillance system, more and more surveillance videos are produced. One has to manage and real-time mine information and knowledge of interest from the large-scale videos, in order to realize intelligent video surveillance. The moving object detection and tracking methods are the most basic and important technology in the area of intelligent video surveillance, and are the key to realizing real-time intelligent video surveillance. The traditional methods, which are not fit for complex environment and can’t meet accuracy and real time at the same time, are common designed for special scenes.Due to the requirements above, this thesis focuses on the rigorous and real-time methods for both detecting and tracking moving objects within video sequences acquired by monocular and static camera. The primary work including:(1) The scene of video surveillance system acquired by static camera is partitioned into two parts: the simple scene and the complex scene, which is based on the complex degree of video background model. A moving object detection method based on area partition for simple scene is proposed and improves the detection accuracy and low illumination sensitivity, also a fast convergent Gaussian Mixture Model for complex scene is proposed and improves the convergent speed and time efficiency compared with the traditional Gaussian Mixture Model.(2) Moving object detection can acquire the binary image of background and objects. A fast path-based binary image clustering method and a fast holes filling method are proposed to carry out objects segmentation and noise elimination. Compared with traditional methods, the proposed methods account for the low and unstable time efficiency and enhance the clustering and filling effect.(3) Most moving objects in surveillance video are human or vehicles. Human’s 3D upright ellipsoid model does well in solving human’s part sheltered and shadow problem, but that vehicle model is used to solve vehicle’s part sheltered and shadow problem will be a time-consuming work because of the complexity of the vehicle model. To this end, in this thesis, a part sheltered vehicle segmentation and shadow elimination method based on Morphology which improves the time efficiency is proposed, and the method is real-time.(4) A moving object tracking method based on adaptive particle filter, which is not dependent on fixed moving model and improves the accuracy, is proposed. A moving object tracking method based on forecasting and character matching, which picks up moving object’s figure, geometry and color characters, and improves the accuracy of particle filter, is proposed. The experimental results corresponding to each method are presented and the efficiencies of the methods are evaluated and discussed under the criterion of accuracy and real time. The researches of this thesis will make contribution to the technology of moving object detection and tracking in surveillance video acquired by static camera.更多还原