【作者】 杨广林；

【导师】 孔令富；

【作者基本信息】 燕山大学 ， 计算机应用技术， 2006， 博士

【摘要】 随着人们安全防范意识的增强,视频监控得到了广泛的普及,越来越多的摄像机被安装在机场、商店、停车场、交通路口等场所。为了提高监控的准确性和监控效率,视频监控需要向智能化方向上发展,产生了所谓的智能视频监控。这一技术包括在视频图像序列中自动地进行运动目标的检测、目标跟踪、目标分类和行为理解等方面的内容,目的是在图像及图像描述之间建立映射关系,从而使计算机能够理解视频画面中的内容。本文将可移动的动态摄像机引入到由静止摄像机构成的监控系统中,讨论该系统所涉及到的摄像机标定、手眼关系的标定、移动目标的提取和跟踪、两个摄像机之间目标的匹配以及对动态摄像机运动平台的伺服控制等问题。首先讨论了摄像机标定问题和手眼关系的标定问题,这方面的具体工作如下:(1)提出的一种利用环境中目标物体运动的已知信息,求解摄像机的内参数问题的方法。该方法的基本思想是:让被摄物体作三个不在同一平面的平移运动,然后,根据每次平移运动前后摄像机获取的两幅不同图像特征点的对应关系,建立方程,求解摄像机的内部参数。通过引入了中间变量,以避免求解非线性方程,在物体上选择多个特征点,可以得到一个线性方程组。并且考虑未知量之间的约束条件,利用Lagrange乘数法对方程组求解。(2)提出了一种内参数异构情况下摄像机平移位置的测定方法。该方法核心是给出了焦距变化时图像平面极点的求解方法,进而可求出摄像机的运动方向,并实现摄像机平移位置的测定,同时也给出了在摄像机焦距调整后,焦距检测的简单方法。(3)提出了一种新的手眼关系的标定方法。该方法是一种自标定方法。需要场景中两个特征点,通过控制摄像机运动平台的做两次平移和两次旋转运动,即可通过图像之间的对应关系求出手眼关系的旋转矩阵和平移向量。与以往算法的不同之处在于,在计算手眼关系的平移向量时,让摄像机平台进行纯旋转运动,再对摄像机坐标系进行虚设旋转变换使旋转转化为平移问题。同时,也给出了基于主动视觉的空间点深度值计算方法。其次,由于复杂的环境因素,使得背景模型越来越复杂。为了提高运动目标检测的执行效率,提出了一种基于图像块的多像素背景模型的构建方法,该方法将视频图像分块,以图像块的特征来构建背景模型。并给出了基于图像块的高斯混合背景模型和LOTS背景模型。并进行了实验验证。最后,为了克服由静止摄像机组成的监控系统对目标跟踪的不足,将手眼系统引入到视频监控中。提出了一种主动的视频监控系统,它是由一个固定的静止摄像机加上一个可移动的动态摄像机组成的双摄像机监控系统。其目的是对异常目标进行实时的跟踪并控制动态摄像机运动平台使所要跟踪移动目标总是出现在动态摄像机图像的中心位置。该系统在标定摄像机和手眼关系的基础上,利用所监控的环境特点,给出两个摄像机图像平面之间的近似的单应性关系,并以此为基础,建立两个摄像机之间目标匹配的方法。系统在静态摄像机的图像平面上建立目标的2D运动模型,采用卡尔曼滤波实现目标运动位置的预测,然后再利用单应性关系,得到在动态摄像机图像平面上对应目标的位置预测,计算动态摄像机平台所要旋转的角度,实现对该摄像机的运动控制,同时也讨论了系统的误差补偿等技术问题。更多还原

【Abstract】 Along with the boost up of people’s security consciousness, video surveillance has become very common. More and more cameras have been installed at airports, stores, parking plots and traffic crossroads etc. In order to improve the veracity and efficiency of surveillance, we should develop video surveillance in the direction of intelligentizing. The technique includes automatically carrying detection of moving objects, objects tracking, objects classifying and behavior understanding in video image sequence, etc. The purpose of the technique is to build a mapping relation between image and description of the image, thereby making computer be able to understand the content of video images. In this paper we introduce a removable camera into surveillance system consisting of static cameras. We discuss some problems related to the system, such as camera calibration, hand-eye relation calibration, moving objects detection and tracking, objects matching between these two cameras and the servo control of a dynamic camera etc.First of all, we discuss the calibration of cameras and eye-hand relation, in which the main work is as the following: (i) We present a method of solving inner parameters of cameras by using known kinematics information about the target objects in surroundings. The main idea of the method is: let the object shot make the translation movements in three directions which are not in the same plane, then by using the corresponding relationship of characteristic points on the two different images before and after the translation movement, we build up the equation and solve the inner parameters of the camera. Introducing the affixation variables instead of solving non linear equation, we can make use of several corresponding points on the object and obtain a group of linear equations. According to the restricting conditions of the variables, we solve the equation by means of Lagrange’s Multiplier; (ii) We also present a method of detecting the translation position of the camera under its different inner parameters at different positions. The core of this method is putting forward a way to solve epipolar of image plane while focus distance varies. Besides, we make sure of the translation direction ofcamera coordinate system and give a method of examining camera translation position too. At the same time two simple methods of examining focus distance after adjusting camera focus distance are given. (iii) We come up with a new hand-eye calibration method, which is a kind of self-calibration method. It needs two characteristic points in the scene and controlling the plate do two translations and two rotations of camera plate to calculate rotation matrix and translation vector of hand-eye relation based on the correspondence between images after the camera movement. The difference from other methods is that as calculating translation vector of hand-eye relation, this one firstly makes the camera plate do pure rotating movement and then utilizes nominal rotation to make rotation change into the translation. At the same time we also give the calculating method for depth value of the points in the scene based on the active vision.The second, complex factors in the surrounding make the background model more and more complicated. In order to improve executing efficiency of examining moving objects, we present a method of constructing multi-pixel background model based on the image block. The method partitions the video image into blocks and constructs the background model according to the characteristics of the blocks. We give Gaussian mixture background model and LOTS background model based on the image block and carry out experiments to validate it.The last, in order to overcome the shortcomings of surveillance which consists of static cameras to track objects, hand-eye system should be introduced into video surveillance. We bring forward a kind of active video surveillance system. The system, composed of two cameras --- one static camera and the other removable camera, is a real time bi-camera tracking system. The aim of the system is to carry out the real time tracking down the unconventional object and to control the locomotion plate of the removable camera so as to make the object be tracked appear in the centre of the image in the dynamic camera. On the basis of calibration of cameras and hand-eye relation of the system, we give approximate homography relation between two cameras’image planes by making use of the characteristics of the surroundings watched. From this we can establish a method to correspond the objects on the two image plates between two cameras. In the system, 2D motion model is established on the image plane of the static camera to predictthe position of the object by Kalman filter. Then by using homography relation, the prediction of the object position can be obtained on the image plane of the dynamic camera. Finally, after calculating the rotation angle of the dynamic camera plate, we can implement the servo control of the dynamic camera. We also discuss the technical problems of the error compensations of the system.更多还原