【作者】 陈功；

【导师】 周荷琴；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 智能视频监控(Intelligent Visual Surveillance,IVS)是计算机视觉领域一个新兴的研究方向,与传统的视频监控系统相比,它不仅能用摄像机代替人眼,还将借助于计算机软件构建一个自动化或半自动化的视频理解和分析系统,以提供及时准确的视频分析结果,并根据需要及时发出报警信息。主要研究内容包括运动物体的检测、运动物体的识别和跟踪、异常现象的检测和报警等。该系统在交通、公安、海关、银行、军事等许多领域都有重要的用途。智能视频监控系统的研究已引起国际上许多著名科研机构和研究人员的兴趣,如美国的Carnegie Mellon大学,国内的中科院自动化所等,都已经实现了自己的智能视频监控平台。为了提高监控系统的可靠性和智能化程度,必须研究鲁棒的视频监控算法,相关的研究涉及到许多难题。本论文对鲁棒的智能视频监控方法进行了深入细致的研究,研究对象是交通路口、住宅小区的监控视频,给出了一个鲁棒的智能视频监控的解决方案,能够适应复杂天气状况下的目标检测和跟踪,包括白天、夜间、雾天的场景,并在监控中完成对阴影和遮挡的处理。本文解决了该领域内的一些难题,例如,特殊天气条件下(如雾天)的图像复原算法,处在夜间场景中的运动物体检测方法,运动物体的阴影检测,监控过程中对运动目标遮挡的处理等。归纳起来本文主要完成了以下有特色的研究工作:(1)提出基于无偏卡尔曼滤波器(Unscented Kalman Filter,UKF)的背景提取方法,构建整体的运动物体检测框架。该方法通过对背景建模,分别从帧间差分和背景差分两个层次综合分析象素值的动态变化特性,然后借助UKF对两个模型参数进行在线更新,实现实时的运动物体分割。(2)雾天场景的能见度很低,为了保证视频监控系统的正常工作,本文提出一种新的基于物理模型的雾天图像复原方法,该方法首先对雾天场景的光学成像建模,然后借助于一张晴天和一张雾天的场景参考图像,计算出场景各点的深度比关系,最后利用深度关系复原雾天图像或雾天视频。(3)针对夜晚环境光线较暗的情况,提出了两种夜间图像增强和运动检测方法。算法能够增强夜间低质量图像的对比度,而且夜间车辆检测的效果也比较令人满意。并将夜间和白天的图像融合,使之包含全面的场景信息,更加适用于人体视觉和机器感知。(4)运动阴影的存在会导致多目标粘连,影响目标的识别与跟踪,本文从阴影的特性出发,提出了有效的解决方法,包括改进的基于特征的方法以及基于边缘特征和角点信息的方法,应用于各监控场景,能够实现运动阴影的检测和滤除。(5)针对目标跟踪中存在的遮挡问题,提出了一种基于运动预测框的目标跟踪算法,将它与基于车辆平行四边形轮廓的遮挡分割方法结合,构建了多车辆目标的实时跟踪系统。更多还原

【Abstract】 Intelligent visual surveillance is a new research area in computer vision. It is quite different from the traditional surveillance systems that it not only replaces the human eyes with camera, but also builds an automatic or semi-automatic video understanding and analysis system in the use of computer software. It can offer accurate analysis result, and announce an alarm for abnormal behaviors. Its mainly research topics include moving object detection, moving object recognition and tracking, abnormal behavior detection and alarm. It can be applied in several areas such as transportation, public security, custom, bank and military affairs.Many famous institution and researchers have shown their interest in intelligent visual surveillance system, like Carnegie Mellon university and CASIA. As a matter of fact, they have built the surveillance platform for it. In order to improve the reliability and intelligence, several key problems should be conquered. In this paper, we briefly researched the robust intelligent visual surveillance methods, and brought out a robust solution which can handle object detection and tracking in bad weather conditions, including daytime, nighttime and foggy day situation, and can also take care the affection of moving shadow and occlusion. We solved several key problems such as image restoration method in foggy day, moving object detection in nighttime, moving shadow detection and tracking with object occlusion. Our meaningful and detailed research work is organized as follows:(1) An Unscented Kalman Filter(UKF) based background subtraction method is proposed, and a whole moving object detection frame is constructed. Background is modeled firstly. Then the dynamic change of pixels is analyzed through two levels which are frame to frame differencing and background differencing. Finally, UKF is used to update the model parameters online, and realize real-time moving object segmentation.(2) Scene visibility is very low in foggy day. We proposed a novel physical model based defog method to make sure the intelligent visual surveillance system work normally. This method models foggy day scene points firstly. Then scene depth is calculated with one clear day image and one foggy day image. The foggy day image or video is restored using depth information finally.(3) According to the condition of nighttime environment, we proposed two novel methods for nighttime video enhancement and moving object detection. They can enhance the original low quality nighttime image, and the experimental results of the moving object detection show that our methods are effective and satisfying. Moreover, the final fused daytime and nighttime image contains a comprehensive description of the scene which is more useful for human vision and machine perception.(4) Moving shadows would connect moving objects together, and affect object recognition and tracking. From the feature of moving shadows, we proposed two novel methods to handle this situation, including improved feature based method, and edge feature and corner point information based method. Applied to different surveillance scenes, our methods can detect and remove moving shadows well.(5) In order to solve the occlusion problem in object tracking, a movement estimation frame based tracking method is proposed, and combined with parallelogram contour based occlusion segmentation method to realize a robust real-time multi-vehicle tracking.更多还原