视频中背景建模目标检测综述(北京大学)
基于视频的移动目标检测是一个重要且有挑战性的任务，在许多应用中都起到相当关键的作用。本次论文研读围绕该主题展开，深入阅读了十余篇论文，在本文总结了视频中移动目标检测的一些主要方法及各自的优劣。报告接下来的部分组织如下:第二节介绍一般问题的陈述及典型的应用,第三节与第四节分别介绍基于像素的移动目标检测方法与基于区域的移动目标检测方法,第五节进行简单的讨论与总结。2问题陈述及应用一般而言移动目标检测并不单独地构成应用,而是作为一个组件出现在许多实际的应用之中。故移动目标检测的具体要求随着应用的改变而有很大的不同。例如对足球场上球员及足球的检测与跟踪就和对视频中用户手势的跟踪有所不同,前者的关键在于如何应对复杂的光照变化有效提取运动物体,后者的难点则在于如何从整个躯休大范围的运动背景中将手势的运动识别并提取出来。尽管不同的应用可能提出不同的技术上的要求,但是相当一部分这类问题还是可以在同一的框架下进行探讨和比较的。以下就是本文讨论范围内一般问题的陈述( Figure1):a.高层次的模型一般具有检测、跟踪、识别三个模块,其中识别模块并不必须b.检测模块可细分为移动日标区域检测与移动目标分组,其中前者是这个模块能够顺利工作的保证,其日的是将各帧内移动日标所在的区域标出该模型的输入为连续的视频,输出为跟踪的物体(即轨迹)或分类的物体,DetectioTrackingbackgroundObject detectionMatching usingFramescolor texture andTrackedsubtractionusing contoursmotion featuresobiectsGr。 up handling(merging andlittonFigure1系统框架图中即为一个典型的流程图,其中省略了分类模块并把运动目标区域检测规定为背景差分方法。该模型的典型应用场景是室内外的视频监视分析( Figure2),特别是交通数据的分析。另外体育运动视频(如足球或台球)的分析也能在该模型下解决。Figure2移动目标检测的典型应用:视频监控本文讨论的就是这样一个系统之中移动目标检测部分内容,并且将重点放在了如何判定移动目标区域的部分。这是这样的系统中的最初的处理,对于之后的处理能否获得有效的信息至关重要。该内容主要涉及两个问题:如何提取运动的前景,及如何建立一个良好的背景模型。后者一般并不是应用任务中所要求完成的,但往往是用以提取运动前景很好辅助工具,将新的一帧“减去”背景即可获得移动前景,故同时具有这两个步骤的方法也被称为“背景差分”,是移动目标检测中的一大类主流方法,本报告中涉及的大部分方法即属于这一类3基于像素的移动目标检测31均值-阈限方法均值-阈限( Figure3)的基本思路是计算每个像素的平均值和标准差作为它的背景模型Mean 2*Avg DiffMean”Mean-2*Avg DiffFigure3均值-國限方法4图中为某视频中单个像素在一定时间内不断更新得到的平均值和平均差值3,该像素处在天空的位置,在一段时问后有人的手挥过该区域,可以看到由于前景目标明显不如背景中的天空明亮,所以可以很容易将其分辨出来实际上均值-阈限方法就是赋予视频中每个像素一个统计上的背景模型,例如高斯分布模型4。每个点需要两个参数来衡量:均值与方差。后面将看到,许多更先进的棊于像素的移动目标检测方法其实无非采用了更复杂的分布模型来描述每个像素32阴影去除及三高斯模型简单的帧差值或均值-阈限方法在很多应川中都面临一个很严重的问题:阴影。在某些光照条件下,移动物体产生的阴影相对背景具有非常显著的差别从而被识别成了前景,有时这些阴影比物体本身还人,并且导致原本独立的运动物体连接在一起无法分割。一种简单的思路是放弃使用灰度值进行背景建模,而采用颜色信息从而将阴影的移动去除掉。这类方法需要一条假设:移动目标投射到路面上的阴影主要改变了该位置的亮度而对色度没有大的影响°。部分情况下确实可以承认该假设。在不同的颜色模型下有不同提取亮度信息的方式。在HSV等空间中这个任务尤为简单,因为亮度本身就是一个独立的分量,所以在该分量以外的维度上进行背景建模与差分就能消除一定的阴影。如果在常用的RGB颜色空间中,亮度的提取就稍微复杂倒。Figure4RGB空间中的亮度与色度3类似标准差的作用,但是计算更快捷。C.R. Wren, A Azarbayejani, T. Darrell, and A P Pentland "pfinder Real-Time Tracking of the human body lEEETrans. Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp. 780-785, July 1997J.M. Alvarez, A Lopez, and r Baldrich, " Illuminant-Invariant Model-Based Road Segmentation",IEEE IntelligentVehicles Symposium, June 20085将一个像素的颜色值在RGB三维空间中表示( Figure4),背景建模就是确定了穿过原点的一条色度直线,所有在该直线上的颜色都认为是背景色。当前颜色相对参考颜色(背景)的亮度分量a1由最小化下式给出:φ(a;)=(l1-c1E)2a2表示该像素当前值相对参考背景色的相对亮度。如果其值为1则代表亮度相同,大于1代表比背景更亮,小于1代表比背景更培当前颜色到色度直线的垂直距离就表小色度的偏离:CD1=‖l2-aE1‖(3)利用色度与亮度的偏离值就可以将新的颜色值分为四个类别Forground: CDi> IcD or ai> Alow, elseM()Back ground:aτa2else(4)ShadowelseHighlightotherwise其中CD2与a分别代表均一化之后的两个偏离分量。这些方强烈依赖于上面提到的假设:阴影仅仅改变背景亮度而不改变色度。但是这个假设实际上并不总是有效,很多时候还需要更复杂的方法达到去除阴影,鉴别真正移动目标的目的。三高斯模型4就是针对去除阴影的考虑提出的。这个模型中采用三个高斯分布相结合( Figure5)对各像素进行建模。三个高斯成分分别为:道路、运动前景及阴影。这三个成分组合成为了完整的混合模型(b)和印甲品f叫intersity valeFigure5三个高斯分布相结合6此方法之后面临的主要问题是如何通过一定时间的学习获得每个高斯分布的参数从而建立有效的模型。相关文献中提出采用EM算法( ExpectationMaximization Algorithm)进行学习。EM算法是一个迭代的算法,通过有限步的迭代就能够获得较好的模型估计。一般而言,为了从一个数据集中获得该数据集满足的混合分布,可以采用最大后验概率估计的方法进行估计,但是这样的方法需要关于每个数据分类的信息(即每个值属于哪个类别)。然而在移动目标检测过程中往往都是无监督的学习从而不可能获得这样的分类信息,而只能自动设定个预先的分类,然后通过迭代不断改进,这就是EM算法的基本思路另外,由于各点的数据是不断改变的,于是采川原始的EM算法对每一帧都进行重复的迭代既不必要也不现实,可以采用EM算法的一个变种:增量EM算法33高斯混合模型(GMM331背景建虞在某些场景之下,采用三个高斯分布的混合模型仍然无法有效地描述复杂的现实环境,于是髙斯混合模型四被提出了。高髙斯混合模型采用类似3.2中三个髙斯模型的思路,希望采用多个高斯分布相结合的方法来描述环境。与前面的模型不同的是,现在高斯分布的个数不是固定的一个或三个了,而是随着各个像素实际的需要动态地进行设定。另外该方法也放弃采用费时的EM算法而采用更快捷的方式进行背景建模与更新。假设已知像素(xo,y)在过去一段时间中的颜色值或灰度值{X1,…,X}={(xo,yo,):1≤i≤t(5)若由K个高斯分布的高斯混合模型对该像素进行建模,则新观察到一个颜色值或灰度值的概率为P(x1)=)o*n(x,,E(6)在RGB等彩色空间中为了简化计算可以采用如下公式k,t7)A Dempster, N. Laird, and D. Rubin, "Maximum Likelihood from Incomplete Data via the EM algorithm", Journalof the Royal Statistical Society, pp. 1-38, 1977这个公式假设了各个颜色分量是相互独立的。尽管事实并不如此,但这样计算在保留充分的精确性的同时大大降低了计算复杂性。这样每个像素就采用κ个峰的髙斯混合模型完成了建模。剩余的问题同样是如何对模型中的各个参数进行估计,以及如何判别前景。3.3.2背景更新背景更新大致有这几个步骤:a.每一个新的值都与所有K个高斯分布进行匹配,直到找到充分吻合的分布(判断方式与均值-阈限方法类似)b.若新的值与所有K个分布都不接近,则将K个分布中权值最小的一个替换成一个新的分布,该分布以新的值为均值,并且具有很高的初始方差和很低的初始权重;C.若新的值与某个分布充分接近,则认为其属于该分布并更新各个分布的权值及参数Ort=(1-aOKt-1+aM(8)=(1-p)ut-1+p·X(9)=(1-p)21+p(X-1)(x-2)(10)here(kok(11)其中(8)式中Mkt对于匹配上的分布唯1,对于其他分布为0α表示适应性的强弱,α越大,给予新的值的权重越大,也就适应得越快;(9)式(10)式仅针对匹配上的分布,而其他分布的参数保持不表3.3.3前景检测运动前景的检测主要有以下几个步骤a.对所有的高斯分布按照ω/σ进行从大到小的排序;b.取前B个高斯分布,满足B= argminkT(12)k=其中T表示背景应该占的比重,如果T取得较小则类似于前面介绍的均值阈限模型,T取得较大则允许背景有更丰富的特性,如随风摆动的树叶或水面的波纹等等。C.如果当前出现的新的值并不符合这B个高斯分布,则认为是运动的前景,否则认为是背景。34非参数模型341背景建模为了能够更快速地适应变化的背景,并且保证对移动物体的敏感性,马里兰大学A. Elgammal等人提岀采用非参数的模型四对各个像素进行建模。该方法并不指定确切模型形式,而釆用核旳数来利用历史薮据建立模型。在釆用高斯核的情况下,一个颜色出现的概率’∑K(x-x∑∏1(13)2这样实际上就是对考虑范围内所有的历史值都建立一个高斯分布,并利用所有这些分布对当前值进行分析。由于高斯分布的假设,故(x1+1-x1)~N(0272),于是可以估计方差(14)0.68V2其中m是|x+1-xl的中位数。号外由于在这种方法下需要大量计算核函数的值,故可以预先计算出一定精度的核函数数据表,通过查表的方法大大加快计算的速度。实验表明这样的方法在一般的PC机上是可以达到实时的计算要求的。34.2减少错误检测为了降低由于局部抖动(如树叶抖动、摄像机抖动)带来的错误检测,该方法利用了一定的区域信息。原先的P(x1)现在改用Pm(x代替PN(t= maxyEN(x)Prix Byy(15)此处N(x)指该像索周围的一个小区域,B指像素y对应的背景模型。这样就充分降低了由于小范围抖动导致的错误检测。7该式成立仍然需要假设各个颜色分量相互独京。9该方法需要维护两个背景模型:长期模型及短期模型。其中短期的模型是为了能够快速适应变化的背景而提出的,仅利用相当短的一段历史值建立模型,另外该模型采用选择更新机制(仅对判定属于背景的值进行更新);长期的模型在相当长的时间内通过盲更新机制(对所有值进行更新)获得。两个模型判定结果的交集能够进一步降低错误的检测,但同时也去除了部分实际上是移动目标的部分。最终采用的策略为:所有由短期模型检测出来,并且与两个模型判定结果交集相邻的像素被视作运动的前景3.4.3去除阴景这个方法同样提到了去除阴影的问题( Figure6)采用RGB颜色模型的一个变种表示颜色:RBR+G+B·9=R+G+Bb=R+G+B′(16)s=RtG+B(17)令A为某个像素在一定时间内的取值,而定义B如下B={x1|x∈A≤≤阝(18)其含义就是该像素过去的背景取值中与当前值亮度接近的值的集合利用B中的值在(r,g)二维空间上进行上述的背景建模及前景检测,就能很有效地消除检测的阴影。Figure6非参数模型下的阴影去除4基于区域的移动目标检测从上面的介绍的方法来看,建立的背景模型越来越复杂:高斯分布的个数从一个到三个,再到K个,再到每个历史值各一个。这样的背景模型已经具有了高度的复杂性,但是在有些应用环境下效果仍然并不理想。反思一下这些方法的特点,仅仅利用各个像素的独立信息而没有考虑像素之间的关联性是键的原因。其实在3.42中采用方法实际上已经开始尝试以局部区域作为考虑问题的范围,
- 2020-12-03下载
- 积分：1
RTCM3.3协议全
全新RTCM3.3协议完整版RTCM STANDARD 10403.3DIFFERENTIAL GNSS(GLOBAL NAVIGATION SATELLITE SYSTEMS)SERVICES – VERSION 3DEVELOPED BYRTCM SPECIAL COMMITTEE NO. 104OCTOBER 7, 2016COPYRIGHT©2016 RTCMRadio Technical Commission for Maritime Services1611 N. Kent St., Suite 605Arlington, Virginia 22209-214RTCM Paper 141-2016-SC104-STD000ocRTCMco00c30RTCM 10403. 3, Differential GNSS Global Navigation Satellite Systems)Services- Version 3, October 7, 2016This standard (referred to as version 3 has been developed by rtCm special Committee 104 as a moreefficient alternative to the standards entitled rtcm recommended standards for diffe rentialRecommended Standards for Differential gNss Global Navigation Satellite Systems Service, Version 2.x(Current version is 2. 3, now designated as RTCM 10402. 3. Service providers and vendors represented onthe SC104 Committee wanted a new standard that would be more efficient, easy to use, and more easilyadaptable to new situations. The main complaint was that the version 2. x parity scheme, which useswords with 24 bits of data followed by 6 bits of parity, was wasteful of bandwidth. Another complaint wasthat the parity was not independent from word to word. Still another was that even with so many bitsdevoted to parity the actual integrity of the message was not as high as it should be. Plus, 30-bit wordsare awkward to handle. the new standard version 3 is intended to correct these weaknessesUnlike Version 2. x, this standard does not include tentative messages The messages in Version 3 haveundergone testing for validity and interoperability and are considered to be permanent. amendments tothe standard may change the meaning of reserved bits or provide additional clarifying text, but no changeswill be made in the data fields. Changes will require new messages to be developed. In addition to themessages described in the current standard the committee continues to develop new messages whichare described in separately published amendments and periodically gathered into a new edition of thestandard. RTCM 10403x for dgNSS services is proving useful in supporting highly accurate differentialand kinematic positioning as well as a wide range of navigation applications worldwideNote that Version 3 messages are not compatible with Version 2. x. Since many receivers have beendesigned and programmed for use with Version 2. x messages, rtCm is maintaining both standards0402 3 and 10403, 3 as" standardsVersion 3.0The initial edition consisted primarily of messages designed to support real-time kinematic (RTK)operations. The reason for this emphasis is that rtk operation involves broadcasting a lot of informationand thus benefits the most from an efficient data format. Version 3.0 provided messages that supportGPS and gloNaSs rTK operations including code and carrier phase observables antenna parametersand ancillary system parametersVersion 3. 1(RTCM Standard 10403.1:The next edition, Version 3. 1 (RTCM Standard 10403. 1), incorporated GPS Network Corrections, whichenable a mobile receiver to obtain accurate rtk information valid over a large area. In addition, new GPSand GLoNaSS messages provide orbital parameters to assist in rapid acquisition a Unicode text messageis also provided for the transmission of textual data. Finally a set of messages are reserved for vendorswho want to encapsulate proprietary data in their broadcasts the gps Network Corrections enable amobile receiver to obtain accurate rtk information valid over a large area. the network rtk correctioninformation provided to a rover can be considered as interpolated corrections between the referencestations in the rtk network this interpolation is not perfect and varies with the actual conditions of theatmosphere. A residual interpolation error has to be expected. With sufficient redundancy in the RtKnetwork, the network server process can provide an estimate for residual interpolation errors. Suchquality estimates may be used by the rover to optimize the performance of rtk solutions The values maybe considered by the rover as a priori estimates only with sufficient tracking data available the rovermight be able to judge residual geometric and ionospheric errors itselfVersion 3. 1. Amendment 1:Amendments 1 was an extensive addition that adds rtcm messages containing transformation data andinformation about Coordinate reference Systems. For rtCm data supporting a rtk service, coordinatesare measured within the itrf or a regional realization surve yors and other users of rtk services mustnormally present their results in the coordinates of local datums. Therefore, coordinate transformationsare necessary. by having RTCM messages that contain transformation data and information about theCoordinate reference systems the users of the rtk service can obtain their results in the desired datumwithout any manual operations. the rtk service providers can then ensure that current information forthe computation of the transformations is always used. the convenience of this method will promote theacceptance of rtK servicesVersion 3. 1. amendment 2:Amendment 2 added residual error messages to support the use of Non-Physical or Computed referenceStations in a network rtk environmentVersion 3. 1. amendment 3:Amendment 3 addressed differences in the way gnss receiver manufacturers have implemented carrierphase encoding of some Version 3 messages so that carrier phase observations are in phase for all carrierphases of a specific frequency i e. they correct for quarter cycle phase shifts. others retain the quartercycle offset between the carrier phase observations in the data. this amendment documents the waydifferent manufacturers have handled the phase shift issue and prescribes a uniform approach for futureproducts.∨ersiⅰon3.1, Amendment4:Amendment 4 added sections 3.5.13 on glONASS Network rtK Correction Messages and 3.5. 14 on FKPNetwork Rtk Correction Messages Related revisions were also made elsewhere in the document.Version 3. 1. amendment 5Amendment 5 added section 3. 5. 12 on State Space Representation related revisions are also madeelsewhere in the document, along with some editorial correctionsVersion 3. 2(RTCM Standard 10403.2)Version 3.2 consolidates Version 3. 1 and all five amendments into a new edition, and it adds MultipleSignal Messages (MSM)as well. the Multiple Signal Message (MSm)format generates receiverobservables in the same way for all included satellite systems. the messages include compact and fullmessages for Pseudorange, PhaseRange, Carrier to Noise Ratio (standard and high resolution), andPhaseRangeratea table near the beginning of the standard lists which messages were included in each separate editionand amendment, so it should not be necessary for users to refer to older versions. Multiple signalMessages are a generic format that will be followed for all GNSs systems. version 3 originally consisted ofmessages for GPS and GLONASS, each in their own format Now with the imminent addition of signals forBeiDou, Galileo, and QZSS, as well as new signals provided by modernized GPS and GloNASS satellitesthe need for a consistent generic format became evident. service providers and users are urged to migrateto the MsM messages to make it easier to accommodate new gNss services(See The RTCM Multiple Signal Messages: A New Step in GNSS Data Standardization")Another newmessage is the gloNaSS Bias Information message. This message provides information which is intendedto compensate for the first-order inter-frequency phase range biases introduced by the reference receivercode- phase biasVersion 3.2, Amendment 1:Added Galileo F/NAv Satellite Ephemeris Data(msg. 1045 )and Bds MSM(msgs. 1121-1127)Version 3.2 amendment 2Added qzss ephemeris(msg. 1044 )and QZss MSm (msgs. 1111-1117Version 3. 3(RTCM Standard 10403.3)This new edition adds Satellite-Based Augmentation System Multiple Signal Messages to previouslydopted messages for GPS, GLONASS, Galileo, and QzssA new ephemeris message has been added for BeiDou(BDS)and a new I/NAV ephemeris message hasbeen added for Galileo. The new edition also reserves 100 messages be used exclusively by sc104 fornew message developmentFinally, the new edition makes consolidates previous amendments and makes numerous editorialImprovementsNavstar GPS Service, Version 2. x. Service providers and vendors represented on the scco000c30z1O2co00c30Contentsco00c30
- 2020-06-27下载
- 积分：1
