• PUDN用户
    了解作者
  • PDF
    开发工具
  • 226KB
    文件大小
  • rar
    文件格式
  • 0
    收藏次数
  • 1 积分
    下载积分
  • 44
    下载次数
  • 2009-11-24 16:46
    上传日期
基于粒子滤波方法实现的实时人脸目标跟踪,程序步骤清楚,方法详细。
RealTimeFaceTrackingUsingPFMEANSHIFT.rar
  • Real Time Face Tracking Using PF MEANSHIFT.pdf
    265.5KB
内容介绍
<html xmlns="http://www.w3.org/1999/xhtml"> <head> <meta charset="utf-8"> <meta name="generator" content="pdf2htmlEX"> <meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1"> <link rel="stylesheet" href="https://static.pudn.com/base/css/base.min.css"> <link rel="stylesheet" href="https://static.pudn.com/base/css/fancy.min.css"> <link rel="stylesheet" href="https://static.pudn.com/prod/directory_preview_static/62608d07090cbf2c4ed5b626/raw.css"> <script src="https://static.pudn.com/base/js/compatibility.min.js"></script> <script src="https://static.pudn.com/base/js/pdf2htmlEX.min.js"></script> <script> try{ pdf2htmlEX.defaultViewer = new pdf2htmlEX.Viewer({}); }catch(e){} </script> <title></title> </head> <body> <div id="sidebar" style="display: none"> <div id="outline"> </div> </div> <div id="pf1" class="pf w0 h0" data-page-no="1"><div class="pc pc1 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/62608d07090cbf2c4ed5b626/bg1.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x1 h2 y2 ff1 fs0 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x2 h2 y3 ff1 fs0 fc0 sc0 ls1 ws0"><span class="fc1 sc0">978-1</span><span class="fc1 sc0">-</span><span class="fc1 sc0">4244</span><span class="fc1 sc0">-2503</span><span class="_ _0"></span><span class="fc1 sc0">-7</span><span class="fc1 sc0">/08/</span><span class="fc1 sc0">$20.</span><span class="_ _0"></span><span class="fc1 sc0">00 </span></div><div class="t m0 x3 h3 y4 ff1 fs1 fc0 sc0 ls0 ws0"><span class="fc1 sc0">&#169;</span><span class="fs0 ls2 ws1"><span class="fc1 sc0"> 2008 IEEE </span></span></div><div class="t m0 x4 h4 y5 ff2 fs2 fc0 sc0 ls0 ws0"><span class="fc1 sc0"> </span></div><div class="t m0 x5 h5 y6 ff2 fs3 fc0 sc0 ls3 ws2"><span class="fc1 sc0">P</span><span class="fc1 sc0">r</span><span class="fc1 sc0">ocee</span><span class="fc1 sc0">din</span><span class="fc1 sc0">g</span><span class="fc1 sc0">s</span><span class="_ _0"></span><span class="fc1 sc0"> o</span><span class="fc1 sc0">f</span><span class="fc1 sc0"> th</span><span class="fc1 sc0">e IEEE </span><span class="_"> </span><span class="fc1 sc0"> </span></div><div class="t m0 x6 h5 y7 ff2 fs3 fc0 sc0 ls4 ws3"><span class="fc1 sc0">In</span><span class="fc1 sc0">tern</span><span class="fc1 sc0">atio</span><span class="fc1 sc0">n</span><span class="fc1 sc0">a</span><span class="fc1 sc0">l Co</span><span class="fc1 sc0">n</span><span class="fc1 sc0">f</span><span class="fc1 sc0">ere</span><span class="fc1 sc0">n</span><span class="fc1 sc0">ce o</span><span class="fc1 sc0">n</span><span class="fc1 sc0"> </span><span class="_ _0"></span><span class="fc1 sc0">Au</span><span class="fc1 sc0">to</span><span class="fc1 sc0">m</span><span class="fc1 sc0">a</span><span class="fc1 sc0">t</span><span class="fc1 sc0">i</span><span class="fc1 sc0">o</span><span class="fc1 sc0">n</span><span class="fc1 sc0"> an</span><span class="fc1 sc0">d</span><span class="fc1 sc0"> </span><span class="fc1 sc0">L</span><span class="fc1 sc0">o</span><span class="fc1 sc0">g</span><span class="fc1 sc0">i</span><span class="fc1 sc0">stic</span><span class="fc1 sc0">s</span><span class="fc1 sc0"> </span></div><div class="t m0 x7 h5 y8 ff2 fs3 fc0 sc0 ls5 ws4"><span class="fc1 sc0">Q</span><span class="fc1 sc0">i</span><span class="fc1 sc0">ng</span><span class="fc1 sc0">dao, C</span><span class="_ _0"></span><span class="fc1 sc0">h</span><span class="fc1 sc0">i</span><span class="fc1 sc0">n</span><span class="fc1 sc0">a S</span><span class="fc1 sc0">e</span><span class="fc1 sc0">pt</span><span class="fc1 sc0">e</span><span class="fc1 sc0">m</span><span class="_ _0"></span><span class="fc1 sc0">ber 2008</span><span class="fc1 sc0"> </span></div><div class="t m0 x8 h6 y9 ff3 fs4 fc0 sc0 ls6 ws5">Real Time Face Tracking Using </div><div class="t m0 x9 h6 ya ff3 fs4 fc0 sc0 ls7 ws6">Particle Filtering and Mean Shift </div><div class="t m0 xa h7 yb ff4 fs5 fc0 sc0 ls0 ws0">Fan</div><div class="c xa yc w2 h8"><div class="t m0 xb h7 yd ff4 fs5 fc0 sc0 ls0 ws0">g</div></div><div class="t m0 xc h7 yb ff4 fs5 fc0 sc0 ls8 ws7"> Xu </div><div class="t m0 xd h9 ye ff4 fs6 fc0 sc0 ls9 ws0">[1][2] </div><div class="t m0 xe h7 yb ff4 fs5 fc0 sc0 lsa ws8">and Jun Chen</div><div class="c xe yc w3 h8"><div class="t m0 xf h7 yd ff4 fs5 fc0 sc0 ls0 ws0">g</div></div><div class="t m0 x10 h9 ye ff4 fs6 fc0 sc0 ls9 ws0">[2]</div><div class="t m0 x11 ha yf ff5 fs5 fc0 sc0 ls0 ws0">Chao Wang</div><div class="t m0 x12 hb y10 ff5 fs6 fc0 sc0 lsb ws0">[1]</div><div class="t m0 xf hc y11 ff6 fs7 fc0 sc0 lsc ws9">[1]Depart<span class="_ _0"></span>ment of Sof<span class="_ _0"></span>tware, School of N<span class="_ _0"></span>ankai Univer<span class="_ _0"></span>sity <span class="_ _1"> </span>[1]Depart<span class="_ _0"></span>ment of Sof<span class="_ _0"></span>tware, School of N<span class="_ _0"></span>ankai Univer<span class="_ _0"></span>sity</div><div class="t m0 x13 hd y12 ff5 fs7 fc0 sc0 lsd wsa">[2]Shenzhen<span class="_ _2"></span> Institute of Advanced Integration Technolo<span class="_ _2"></span>gy, </div><div class="t m0 x14 he y13 ff3 fs8 fc0 sc0 lse ws0">wangchao@nankai.<span class="_ _2"></span>edu.cn</div><div class="t m0 xf hc y14 ff6 fs7 fc0 sc0 lsf wsb">Chinese Academy of Sciences/<span class="_ _0"></span>The Chinese University of </div><div class="t m0 xd hd y15 ff5 fs7 fc0 sc0 ls10 wsc">Hong Kong</div><div class="t m0 x1 he y16 ff3 fs8 fc0 sc0 ls11 wsd">{fang.xu, jun.cheng}@siat.ac.cn</div><div class="t m0 x4 hf y17 ff7 fs9 fc0 sc0 ls12 wse">Abstract- Particle fil<span class="_ _0"></span>ter is widely<span class="_ _0"></span> used in ob<span class="_ _0"></span>ject tracking<span class="_ _0"></span>. Howeve<span class="_ _0"></span>r, </div><div class="t m0 x4 hf y18 ff7 fs9 fc0 sc0 ls13 wsf">it has one no<span class="_ _2"></span>table weakn<span class="_ _2"></span>esses that is sam<span class="_ _2"></span>ple degeneracy p<span class="_ _2"></span>roblem. </div><div class="t m0 x4 hf y19 ff7 fs9 fc0 sc0 ls14 ws10">This pape<span class="_ _0"></span>r pro<span class="_ _0"></span>poses a n<span class="_ _0"></span>ovel al<span class="_ _0"></span>gorith<span class="_ _0"></span>m to ove<span class="_ _0"></span>rcome this<span class="_ _0"></span> proble<span class="_ _0"></span>m by </div><div class="t m0 x4 hf y1a ff7 fs9 fc0 sc0 ls15 ws11">incorporatin<span class="_ _0"></span>g mean shif<span class="_ _0"></span>t into pa<span class="_ _0"></span>rticle filtering. Me<span class="_ _0"></span>an shift reacti<span class="_ _0"></span>ng </div><div class="t m0 x4 hf y1b ff7 fs9 fc0 sc0 ls16 ws12">on samp<span class="_ _2"></span>le herds the samp<span class="_ _2"></span>les in the reference m<span class="_ _2"></span>od<span class="_ _2"></span>e area, which </div><div class="t m0 x4 hf y1c ff7 fs9 fc0 sc0 ls17 ws13">could make le<span class="_ _0"></span>ss samples be u<span class="_ _0"></span>sed while trackin<span class="_ _0"></span>g. The propo<span class="_ _0"></span>sed </div><div class="t m0 x4 hf y1d ff7 fs9 fc0 sc0 ls12 ws14">algorithm is<span class="_ _0"></span> used in fac<span class="_ _0"></span>e track<span class="_ _0"></span>ing. Results de<span class="_ _0"></span>monstrate that <span class="_ _0"></span>our </div><div class="t m0 x4 hf y1e ff7 fs9 fc0 sc0 ls18 ws15">approac<span class="_ _0"></span>h has<span class="_ _0"></span> better<span class="_ _0"></span> perfor<span class="_ _0"></span>mance tha<span class="_ _0"></span>n that<span class="_ _0"></span> of the<span class="_ _0"></span> mean<span class="_ _0"></span> shift<span class="_ _0"></span> </div><div class="t m0 x4 hf y1f ff7 fs9 fc0 sc0 ls19 ws16">tracker and t<span class="_ _0"></span>he conventiona<span class="_ _0"></span>l particle filter. Moreove<span class="_ _0"></span>r, the </div><div class="t m0 x4 hf y20 ff7 fs9 fc0 sc0 ls1a ws17">computation<span class="_ _0"></span> time in each fr<span class="_ _0"></span>ame is less than <span class="_ _0"></span>that of the<span class="_ _0"></span> mean shift </div><div class="t m0 x4 hf y21 ff7 fs9 fc0 sc0 ls1b ws18">tracker or the c<span class="_ _0"></span>onventional p<span class="_ _0"></span>article filter. </div><div class="t m0 x4 h10 y22 ff8 fs9 fc0 sc0 ls1c ws19">Keywords<span class="_ _0"></span>- particle filte<span class="_ _0"></span>r mean s<span class="_ _0"></span>hift embed s<span class="_ _0"></span>amples </div><div class="t m0 x15 h11 y23 ff3 fs7 fc0 sc0 ls1d ws0">I.<span class="_ _3"> </span>I<span class="fsa ls1e">NTRODUCTION</span></div><div class="t m0 x16 h11 y24 ff3 fs7 fc0 sc0 ls1f ws1a">With the rapid enhancement of computational power, </div><div class="t m0 x4 h11 y25 ff3 fs7 fc0 sc0 ls1f ws1b">object tracking h<span class="_ _0"></span>as been an active research area in the </div><div class="t m0 x4 h11 y26 ff3 fs7 fc0 sc0 ls20 ws1c">computer vision in recen<span class="_ _0"></span>t <span class="_ _2"></span>years. It plays<span class="_ _0"></span> an important role in </div><div class="t m0 x4 h11 y27 ff3 fs7 fc0 sc0 ls20 ws1d">robot tracking<span class="_ _0"></span>, monitoring and surveillance systems, virtual </div><div class="t m0 x4 h11 y28 ff3 fs7 fc0 sc0 ls21 ws1e">sport, hum<span class="_ _0"></span>an-computer interfaces, and etc. </div><div class="t m0 x16 h11 y29 ff3 fs7 fc0 sc0 ls22 ws1f">Three comm<span class="_ _0"></span>on approaches are widely used in<span class="_ _0"></span> object </div><div class="t m0 x4 h11 y2a ff3 fs7 fc0 sc0 ls1f ws20">tracking. One is based on Kalman f<span class="_ _0"></span>ilter or extended Kalman </div><div class="t m0 x4 h11 y2b ff3 fs7 fc0 sc0 ls1f ws1a">filter [13,14]. Som<span class="_ _0"></span>e researchers use mean shift tracker or </div><div class="t m0 x4 h11 y2c ff3 fs7 fc0 sc0 ls23 ws21">snake as the comparison model. Kal<span class="_ _2"></span>man filter assumes that the </div><div class="t m0 x4 h11 y2d ff3 fs7 fc0 sc0 ls24 ws22">probabilit<span class="_ _2"></span>y distributions of the states are Gaussian. However, </div><div class="t m0 x4 h11 y2e ff3 fs7 fc0 sc0 ls25 ws23">in real tracking, many<span class="_ _0"></span> factors lead the tracking states to </div><div class="t m0 x4 h11 y2f ff3 fs7 fc0 sc0 ls26 ws24">distribu<span class="_ _0"></span>te non-Gaussi<span class="_ _0"></span>an and non-lineari<span class="_ _0"></span>ty. Another is mean </div><div class="t m0 x4 h11 y30 ff3 fs7 fc0 sc0 ls20 ws25">shift [2, 4, 16] w<span class="_ _0"></span>hich is a robust non-parametric m<span class="_ _0"></span>ethod for </div><div class="t m0 x4 h11 y31 ff3 fs7 fc0 sc0 ls27 ws26">climbi<span class="_ _2"></span>ng den<span class="_ _2"></span>sity gr<span class="_ _2"></span>adients t<span class="_ _2"></span>o find<span class="_ _2"></span> the pea<span class="_ _2"></span>k of pr<span class="_ _2"></span>obabilit<span class="_ _2"></span>y </div><div class="t m0 x4 h11 y32 ff3 fs7 fc0 sc0 ls28 ws27">model. Mean-Shif<span class="_ _0"></span>t tracker is a real-time algorithm that </div><div class="t m0 x4 h11 y33 ff3 fs7 fc0 sc0 ls1f ws28">endeavors to m<span class="_ _0"></span>aximize the correlation between two st<span class="_ _0"></span>atistical </div><div class="t m0 x4 h11 y34 ff3 fs7 fc0 sc0 lsf ws29">distributions<span class="_ _0"></span>. The correlation, <span class="ls29 ws2a">or sim<span class="_ _0"></span>ilarity between two </span></div><div class="t m0 x4 h11 y35 ff3 fs7 fc0 sc0 ls2a ws2b">distributions<span class="_ _0"></span> is expressed as a measurement deriv<span class="_ _0"></span>ed from the </div><div class="t m0 x4 h11 y36 ff3 fs7 fc0 sc0 ls2b ws2c">Bhattacharyya coefficient. Gen<span class="_ _0"></span>eral distributions may u<span class="_ _0"></span>se </div><div class="t m0 x4 h11 y37 ff3 fs7 fc0 sc0 ls2c ws2d">color, or texture or combine both. But sometimes it falls into </div><div class="t m0 x4 h11 y38 ff3 fs7 fc0 sc0 ls2d ws2e">local extrem<span class="_ _0"></span>e value so w<span class="_ _0"></span>e can n<span class="_ _0"></span>ot find what w<span class="_ _0"></span>e need. Third, </div><div class="t m0 x4 h11 y39 ff3 fs7 fc0 sc0 ls20 ws2f">particle filter [1, 3, 5, 8],<span class="_ _0"></span> as a widely used approach in<span class="_ _0"></span> object </div><div class="t m0 x4 h11 y3a ff3 fs7 fc0 sc0 ls28 ws30">tracking, is well suited to estimate state<span class="_ _2"></span> of non-linear, non-</div><div class="t m0 x4 h11 y3b ff3 fs7 fc0 sc0 ls2e ws31">Gaussian dy<span class="_ _0"></span>namic systems<span class="_ _0"></span>. A common problem<span class="_ _0"></span> with particl<span class="_ _0"></span>e </div><div class="t m0 x4 h11 y3c ff3 fs7 fc0 sc0 ls20 ws32">filtering techniques is the degeneracy problem [3], wh<span class="_ _0"></span>ich </div><div class="t m0 x4 h11 y3d ff3 fs7 fc0 sc0 ls2f ws33">cause the distribution concentrating on a relatively small </div><div class="t m0 x4 h11 y3e ff3 fs7 fc0 sc0 ls30 ws34">subset of sa<span class="_ _2"></span>mple<span class="_ _2"></span>s, thus <span class="_ _2"></span>making t<span class="_ _2"></span>he use<span class="_ _2"></span>ful info<span class="_ _2"></span>rmatio<span class="_ _2"></span>n to be </div><div class="t m0 x4 h11 y3f ff3 fs7 fc0 sc0 ls2a ws35">carried by<span class="_ _0"></span> only a sm<span class="_ _0"></span>all subset of<span class="_ _0"></span> samples wi<span class="_ _0"></span>th significant </div><div class="t m0 x4 h11 y40 ff3 fs7 fc0 sc0 ls2b ws36">weights. To solve this prob<span class="_ _0"></span>lem, there are three k<span class="_ _0"></span>inds of </div><div class="t m0 x4 h11 y41 ff3 fs7 fc0 sc0 ls31 ws37">methods to ov<span class="_ _0"></span>ercome the degen<span class="_ _0"></span>eracy problem: the first is </div><div class="t m0 x4 h11 y42 ff3 fs7 fc0 sc0 ls32 ws38">resampling which is widely used [6, 12<span class="_ _0"></span>]. It selects the large </div><div class="t m0 x17 h11 y43 ff3 fs7 fc0 sc0 ls25 ws39">weights samples and eliminates the small weights ones. It can </div><div class="t m0 x17 h11 y44 ff3 fs7 fc0 sc0 ls21 ws3a">reduce the ef<span class="_ _0"></span>fects of degen<span class="_ _0"></span>erac<span class="_ _2"></span>y phen<span class="_ _0"></span>omenon but introduces </div><div class="t m0 x17 h11 y45 ff3 fs7 fc0 sc0 ls21 ws3b">the samples poor pro<span class="_ _0"></span>blems as well. The s<span class="_ _0"></span>econd one is choice </div><div class="t m0 x17 h11 y46 ff3 fs7 fc0 sc0 ls22 ws3c">of importance dens<span class="_ _0"></span>ity. The importance density is proposed as<span class="_ _0"></span> </div><div class="t m0 x17 h11 y47 ff3 fs7 fc0 sc0 ls33 ws3d">a mixture den<span class="_ _0"></span>sity that depends upon th<span class="_ _0"></span>e past state and the </div><div class="t m0 x17 h11 y48 ff3 fs7 fc0 sc0 ls34 ws3e">most<span class="_ _0"></span> recent observations<span class="_ _0"></span> [17]. Third, em<span class="_ _0"></span>bedding mean sh<span class="_ _0"></span>ift or </div><div class="t m0 x17 h11 y49 ff3 fs7 fc0 sc0 ls33 ws3f">camshift in the particle filter is proposed by Shen and Wei [6]. </div><div class="t m0 x17 h11 y4a ff3 fs7 fc0 sc0 ls2d ws40">They advanced MSEPF(Mean Shift Embedded Particle Filter) </div><div class="t m0 x17 h11 y4b ff3 fs7 fc0 sc0 ls20 ws41">to herd sam<span class="_ _0"></span>ples <span class="_ _2"></span>using mean shift s<span class="_ _0"></span>earch, so it can concentrates </div><div class="t m0 x17 h11 y4c ff3 fs7 fc0 sc0 ls35 ws42">on samples with large weights. The MSEPF can eff<span class="_ _0"></span>icientl<span class="_ _2"></span>y </div><div class="t m0 x17 h11 y4d ff3 fs7 fc0 sc0 ls36 ws43">maintain multiple <span class="_ _2"></span>modes in the posterio<span class="_ _2"></span>r density. </div><div class="t m0 x18 h11 y4e ff3 fs7 fc0 sc0 ls22 ws44">The proposed algorithm<span class="_ _0"></span> is also based on the mean sh<span class="_ _0"></span>ift </div><div class="t m0 x17 h11 y4f ff3 fs7 fc0 sc0 ls37 ws45">and particle filter which ov<span class="_ _0"></span>erco<span class="_ _2"></span>mes the degen<span class="_ _0"></span>eracy problem </div><div class="t m0 x17 h11 y50 ff3 fs7 fc0 sc0 ls1f ws46">effectively. This alg<span class="_ _0"></span>orithm absorbs th<span class="_ _0"></span>e advantages of th<span class="_ _0"></span>e mean </div><div class="t m0 x17 h11 y51 ff3 fs7 fc0 sc0 ls24 ws47">shift and particle filter. In this algorithm, samples are </div><div class="t m0 x17 h11 y52 ff3 fs7 fc0 sc0 ls2c ws48">distributed based o<span class="_ _2"></span>n the po<span class="_ _2"></span>sterior prob<span class="_ _2"></span>ability density, this </div><div class="t m0 x17 h11 y53 ff3 fs7 fc0 sc0 ls38 ws49">means so<span class="_ _2"></span>me sa<span class="_ _2"></span>mples <span class="_ _2"></span>which <span class="_ _2"></span>ma<span class="_ _2"></span>y have a long d<span class="_ _2"></span>istance<span class="_ _2"></span> fro<span class="_ _2"></span>m the<span class="_ _2"></span> </div><div class="t m0 x17 h11 y54 ff3 fs7 fc0 sc0 ls37 ws45">reference m<span class="_ _0"></span>ode area. We use mean shift to s<span class="_ _0"></span>earch the most </div><div class="t m0 x17 h11 y55 ff3 fs7 fc0 sc0 ls25 ws4a">similar area in each s<span class="_ _0"></span>amples and renew each samples center </div><div class="t m0 x17 h11 y56 ff3 fs7 fc0 sc0 ls39 ws4b">with mean shift search. Most sa<span class="_ _2"></span>mples in this algorithm gat<span class="_ _2"></span>her </div><div class="t m0 x17 h11 y57 ff3 fs7 fc0 sc0 ls3a ws4c">to modes of ob<span class="_ _2"></span>servation, so<span class="_ _2"></span> most sa<span class="_ _2"></span>mples have lar<span class="_ _2"></span>ge weights.<span class="_ _2"></span> </div><div class="t m0 x17 h11 y58 ff3 fs7 fc0 sc0 ls3b ws4d">It solves the problems of particle filtering and mean shift </div><div class="t m0 x17 h11 y59 ff3 fs7 fc0 sc0 ls2d ws4e">effectively. <span class="_ _0"></span>Because most s<span class="_ _0"></span>amples have large w<span class="_ _0"></span>eight, it can </div><div class="t m0 x17 h11 y5a ff3 fs7 fc0 sc0 lsc ws4f">hold m<span class="_ _0"></span>odes divers<span class="_ _0"></span>iform in the poste<span class="_ _0"></span>rior density<span class="_ _0"></span> using<span class="_ _0"></span> fewer </div><div class="t m0 x17 h11 y5b ff3 fs7 fc0 sc0 ls2c ws50">samples, resultin<span class="_ _2"></span>g in low computa<span class="_ _2"></span>tional cost. Gao and Y<span class="_ _2"></span>ang </div><div class="t m0 x17 h11 y5c ff3 fs7 fc0 sc0 ls31 ws51">[7] proposed th<span class="_ _0"></span>at using mean shift to find the roug<span class="_ _0"></span>hly area in </div><div class="t m0 x17 h11 y5d ff3 fs7 fc0 sc0 ls2d ws52">the frame, using particle filter in this area. However, the </div><div class="t m0 x17 h11 y5e ff3 fs7 fc0 sc0 ls35 ws53">searching area is large so that it need much ti<span class="_ _2"></span>me. Thus this </div><div class="t m0 x17 h11 y5f ff3 fs7 fc0 sc0 ls20 ws54">approach can n<span class="_ _0"></span>ot solve the dra<span class="_ _2"></span>wback<span class="_ _0"></span> of mean shift, it also </div><div class="t m0 x17 h11 y60 ff3 fs7 fc0 sc0 ls2b ws55">involved in local extreme<span class="_ _0"></span> valu<span class="_ _2"></span>e. </div><div class="t m0 x18 h11 y61 ff3 fs7 fc0 sc0 ls33 ws56">The paper is organ<span class="_ _0"></span>ized as follows. S<span class="_ _0"></span>ection 2 introduces </div><div class="t m0 x17 h11 y62 ff3 fs7 fc0 sc0 ls25 ws57">the particle filter, m<span class="_ _0"></span>ean shift and new<span class="_ _0"></span> algorithm. In section 3, </div><div class="t m0 x17 h11 y63 ff3 fs7 fc0 sc0 lsf ws58">we apply<span class="_ _0"></span> ne<span class="_ _2"></span>w algorithm to face track<span class="_ _0"></span>ing, respectively </div><div class="t m0 x17 h11 y64 ff3 fs7 fc0 sc0 ls27 ws59">discus<span class="_ _2"></span>sing d<span class="_ _2"></span>yna<span class="_ _2"></span>mical <span class="_ _2"></span>model,<span class="_ _2"></span> obser<span class="_ _2"></span>vation <span class="_ _2"></span>model, etc<span class="_ _2"></span>. In </div><div class="t m0 x17 h11 y65 ff3 fs7 fc0 sc0 ls3c ws5a">section 4, we dis<span class="_ _0"></span>cuss our experim<span class="_ _0"></span>ent and compare propos<span class="_ _0"></span>ed </div><div class="t m0 x17 h11 y66 ff3 fs7 fc0 sc0 ls35 ws5b">algorithm with mean shift and conventional particle filter. </div><div class="t m0 x17 h11 y67 ff3 fs7 fc0 sc0 ls27 ws5c">Concl<span class="_ _2"></span>usion a<span class="_ _2"></span>nd fut<span class="_ _2"></span>ure <span class="_ _2"></span>works wi<span class="_ _2"></span>ll be gi<span class="_ _2"></span>ven i<span class="_ _2"></span>n sectio<span class="_ _2"></span>n 5. </div><div class="t m0 x19 h11 y68 ff3 fs7 fc0 sc0 ls1d ws0">II.</div><div class="t m0 x1a h12 y69 ff4 fs7 fc0 sc0 ls0 ws0">P</div><div class="t m0 x1b h12 y6a ff4 fsa fc0 sc0 ls3d ws0">ARTICLE<span class="fs7 ls3e"> F</span><span class="ls3f">ILTER<span class="fs7 ls3e"> W</span>ITH<span class="fs7 ls3e"> M</span><span class="ls40">EAN<span class="fs7 ls3e"> S</span><span class="ls41">HIFT<span class="_ _2"></span> </span></span></span></div><div class="t m0 x17 hd y6b ff5 fs7 fc0 sc0 ls32 ws5d">A. Particle Filter </div><div class="t m0 x18 h11 y6c ff3 fs7 fc0 sc0 ls42 ws5e">Particle filters are a sample-<span class="_ _0"></span>based variant of Bayes fi<span class="_ _0"></span>lters. </div><div class="t m0 x17 h11 y6d ff3 fs7 fc0 sc0 ls43 ws5f">The aim of particle filter estimation is to evaluate the posterior </div><div class="t m0 x17 h11 y6e ff3 fs7 fc0 sc0 ls44 ws60">probabilit<span class="_ _2"></span>y density function, </div><div class="t m0 x1c h13 y6f ff3 fsb fc0 sc0 ls0 ws0">)<span class="_ _4"></span>|<span class="_ _5"></span>(</div><div class="t m0 x1d h14 y70 ff6 fsc fc0 sc0 ls0 ws0">k<span class="_ _6"></span>k</div><div class="t m0 x1e h15 y6f ff6 fsb fc0 sc0 ls0 ws0">z<span class="_ _7"></span>x<span class="_ _8"></span>p</div><div class="t m0 x1f h11 y71 ff3 fs7 fc0 sc0 ls1f ws61"> <span class="_ _9"></span>of th<span class="_ _0"></span>e state vector </div><div class="t m0 x20 h16 y72 ff3 fsd fc0 sc0 ls0 ws0">2252</div><div class="t m0 x21 h17 y73 ff9 fse fc0 sc0 ls0 ws0">Authorized licensed use limited to: TIANJIN UNIVERSITY OF TECHNOLOGY. Downloaded on November 10, 2008 at 21:02 from IEEE Xplore. Restrictions apply.</div></div><div class="pi" data-data='{"ctm":[1.568627,0.000000,0.000000,1.568627,0.000000,0.000000]}'></div></div> </body> </html>
评论
    相关推荐
    • FaceDetectionandHeadTrackinginVideo.rar
      清华大学学位论文;将树状多视角人脸检测器引入粒子滤波器的概率跟踪框架,提出了一种融合了检测信息的跟踪方法;对粒子滤波器的基本算法作扩展,提出了一种融合不同状态量的多个观测模型的方法
    • PF_Video_EN.zip
      粒子滤波。可用于天气预测、故障预测、人脸视频追踪等。
    • paaaabzipa.rar
      粒子滤波算法讲述粒子滤波过程。一种基于Mean Shift算法与粒子滤波器的人眼跟踪,多谢支持。
    • gennou.zip
      多目标跟踪粒子滤波器,Gabor小波变换与PCA的人脸识别代码,合成孔径雷达(SAR)目标成像仿真。
    • 目标跟踪专题.rar
      整理的目标跟踪相关课题,对目前用于的方向进行整理,可作为入门
    • funning.zip
      多目标跟踪粒子滤波器,Gabor小波变换与PCA的人脸识别代码,一种基于多文档得图像合并技术。
    • koupan.zip
      Gabor小波变换与PCA的人脸识别代码,LDPC码的完整的编译码,多目标跟踪粒子滤波器。
    • 目标跟踪综述及相关滤波跟踪论文
      本资源包含目标跟踪的领域较为经典的综述文章,以及相关滤波类目标跟踪算法的论文,包含MOSSE目标跟踪,KCF目标跟踪以及CSR-DCF目标跟踪算法。
    • Face-Tracking-PF-master.zip
      MATLAB仿真工具仿真的face目标跟踪,用粒子滤波对人脸进行跟踪,用于视频中的人脸跟踪
    • 深度学习在视频目标跟踪中的应用进展与展望.rar
      重点介绍了目前发展迅猛的深度学习方法, 包括堆叠自编码器、卷积神经网络等在视频目标跟踪中的最新具体应用情况并进行了深入分析与总结. 最后对深度学习方法在视频目标跟踪中的未来应用与发展方向进行了展望.