节点文献

基于变化区域检测的视频对象分割

Video Object Segmentation Based on the Changing Area Detection

【作者】 陶韬

【导师】 王珂;

【作者基本信息】 吉林大学 , 信号与信息处理, 2004, 硕士

【摘要】 视频对象分割就是把视频按照一定标准分割成区域,目的是为了从图像序列中标记出有意义的实体,这些有意义的实体,在视频中称为视频对象。通常,一个视频场景由视频对象(运动对象)和背景组成,人可以轻易的从视频中分辨出这些部分,如果用计算机来做就很困难。视频中的运动目标是视频的重要组成部分,能够分离出这些部分,在很多领域有着重要作用。在智能监控的应用中,需要实现交通路口、高速公路、机关门卫过往车辆的实时登记和流量统计,这些应用,需要建立在车辆独立分割的前提下;在网络传输中,如果把兴趣对象(视频中一般为运动对象)从需要传输的帧序列中提取出来,分配更多的资源,当信道质量下降时,虽然画面质量跟随下降,但是人们的兴趣中心得以保障,提高了视频的主观质量;在军事领域中,图像往往是重要信息的载体,通过连续拍摄的图像构成序列,从中分离出运动物体能够给侦察提供线索。目前,对红外序列图像的研究以及通过对目标的自动跟踪能够为自动武器的发展奠定基础;在视频编码中,基于内容的编码标准 MPEG-4 已经于 1999 年公布,MPEG-4 采用了基于对象和模型的编码方法,提高了压缩效率,更加合理的分配和利用传输资源。基于对象编码,首先就要对视频对象分割,这个工具并没有包含在标准中,作为标准的开放部分留待进一步研究。就目前的计算能力和图像的复杂性,理想的在任意场景下分离所需对象几乎是不可实现的。针对不同的实际应用有很多适用的方案,但都有适用范围,比较常见的处理方法是根据应用特点,利用视频的一种或者几种特性,结合不同方法的适用范围选择算法实现预定目标。交通领域视频特点是,运动目标为刚体,运动轨迹大多趋于直线。目前在高密度场景的分割很难实现,或效果很难让人满意。门禁、高速公路的监控中,一般目标数目较少,移动速度快。本文的研究目标是针对交通领域、智能监控应用的假设,在摄像机镜头没有移动、变焦等变化,图像背景相对简单、稳定时,单一运动目标乃至于具有相同特征的多运动目标的分割提取。 i<WP=69>吉林大学硕士学位论文在综述目前现存的视频对象分割方案的前提下发现,光流估算的方法运算量大,而且本文研究背景不符合光照不变假设,因此并不适用本文情况。基于块的匹配方法使用灰度作为匹配准则的数据来源,在光照存在突变的区域,基于块的匹配方法也无法正确跟踪目标。本文采用变化区域检测的方法寻找目标初始位置,然后应用基于变形模板的跟踪方法获取目标最终边界。 对于慢速运动物体,可以使用帧间差分的方法提取变化区域,然后进行跟踪修正。但目标运动速度较快的时候,帧间差分的误差很大,差分图像很难体现目标的形状及运动信息。这种情况,可以使用运动目标与背景的差分(减背景法)来检测运动区域。在无法适时直接获取无目标背景,或背景缓变的情况下,背景重建也不得不加入到视频分割的步骤中来。 文中通过分析目前常用的背景重建方法,并针对智能监控应用的需求,提出了更加适合的背景重建方案——基于距离测度的去除干扰法。通过实验比较,Kalman 渐消滤波法基于渐消原理,获得背景后,当运动目标再次扰动时,需要一段时间来消除影响,体现在重建的背景上就是目标后比较长的“拖尾”;基于高阶统计量的分块方法,对于不同的场景,存在恢复失败的可能,致使系统不够鲁棒;本文方法需要较少帧来恢复背景,不同于现存方法的是,本文方法不易受运动目标扰动的影响,不会出现重建失败,而且克服了“块效应”,是适合本文研究背景的重建方法。 基于距离测度的去除干扰法的原理是:对于单个象素而言,运动目标所在区域的象素可以看作是背景象素上的短时扰动。可以把n帧图像的同一位置象素作为一维输入信号做统计,对于偏离中心大的点,判定为车辆的扰动,否则认定为背景象素。在得到n帧图像平均的基础上,依照小流量、高速运动假设,每象素的均值可认为是趋于背景灰度。然后计算每帧图像象素的中心距,中心距大过某预定值,既标定为运动目标象素,恢复背景时去除这些象素的干扰。然后求得更趋近真值的背景灰度。 图像所提供的仅仅是灰度、色彩、纹理、边缘等低等次信息。通常,视频中运动目标带有多种灰度、色彩和纹理信息,通过计算来从中恢复高层次的语义信息比较困难,因此,根据运动特征,运动一致物体的边缘往往能够带来更全面的信息。在众多的分割方法中,Snake 模型具有开放性,用参数可调的方法,动态逼进特征边缘,是一种局部边缘提取的重要方法,因此在 ii<WP=70>摘要图像处理、运动跟踪等方面得到广泛的应用。GVF Snake 模型由 Chenyang Xu在 1997 年提出,它保留 Snake 模型的内部力,通过改进外部能量,使 Snake模型有更大的搜索范围,并且依据图像力判定收缩或者膨胀,不需要事先指定 Snake 的运动方向。同时,解决了 Snake 模型必须初始化在期望边界附近的缺点。 本文仿真实验首先通过去除干扰的方法重建序列图像的背景,应用减背景法获得

【Abstract】 Video object segmentation is to segment video according to certainstandard in order to mark the important objects, which are called video object.General speaking, a video scene is composed of video objects (moving object)and it’s background. It is easy for us to distinguish them, but it is difficult forcomputer to do this. It will be a great merit in many fields if these movingobjects can be divided from the background, for these objects are significantparts in video. In intelligent surveillance, the vehicles passing from the traffic across,highway and guard need to be registered and accounted, which could beaccomplished only when the vehicles are segmented firstly. In internet transfer,if the moving objects can be divided and assigned more resources, when thequality of the picture will decrease, observers interest center can be ensured sothat the subjective quality of video can be increased. In military field, picturesare carriers of important information. Taking successive pictures and dividingmoving objects from it can provide clue for scout. At the present, the researchof infrared sequential images and automatic tracking of objects have laid afoundation for the development of automatic armaments. In video coding, thecode standard MPEG-4 based on content was publicized in 1994.MPEG-4employed the code method based on objects and models, which can increasecompression efficiency and assign transfer resources more reasonably. Codingmethod based on objects requires the video object to be segmented in the firstplace. The instrument is not included in the standard and waits to be studiedfurther. Segmenting objects under any condition ideally can hardly be realized dueto the present capability of calculation and the complexity of images. Eachproject has its own applicability. The usual practice is to choose applicable iv<WP=72>method according to different application characteristics, and making use ofcertain characteristics of video. The video characteristic of traffic field is that moving objects are rigidbodies and the moving tracks are almost straight so that the segmentation inhigh density scenes can hardly be realized satisfactorily. Usually, in thesurveillance of guard and highway, the number of objects is few and the speedsare high. This paper intends to study in the field of traffic intelligentsurveillance how to segment single moving object or multi-objects of the samecharacteristics in a comparatively simple and stable background when thecamera doesn’t move or change focus. When analyzing the existing projects ofvideo segmentation, it can be found that optical-flow estimation cannot beapplied in this paper for it demands a large amount of calculation and theresearch background of this paper isn’t in accordance with the hypothesis oflight invariableness. Since the method of matching block regards gray-level asthe data resources of matching principle, in the region where light breaks, thismethod cannot follow object correctly. This paper searches the original positionof objects by the means of varying region detection, and then gets the accurateboundary of objects through the tracking method based on distortion template. In regard to the low-speed moving objects, variation regions can beextracted by the difference between two frames, and then be traced andmodified. But for the high-speed moving objects, the errors of differencebetween two frames will be great, and the difference image can hardly reflectthe figure and movement information. Under this circumstance, the differencebetween frame and background can be employed to detect the movementregion. When the no-object background cannot be achieved directly in time, orthe background is changing slowly, background rebuilding will have to beadded into steps of video segmentation. Through analyzing the common methods of background rebuilding, andtaking the requirement of intelligent surveillance

  • 【网络出版投稿人】 吉林大学
  • 【网络出版年期】2004年 04期
  • 【分类号】TP274.4
  • 【被引频次】1
  • 【下载频次】332
节点文献中: 

本文链接的文献网络图示:

本文的引文网络