文档名：结合多路梯度与重排序的ADCensus立体匹配算法
摘要：针对局部立体匹配算法在深度不连续和弱纹理区域匹配精度低,且容易受到噪声干扰的问题,提出一种结合多路梯度与重排序的AD-Census立体匹配算法.对传统的AD-Census变换进行改进,重排序不同尺度变换窗口中的像素,取中值作为中心像素点计算Hamming距离,使其对噪声具有更强的鲁棒性,并将八方向梯度信息相融合进行代价计算以增强可靠性;采用十字交叉域和4路径扫描线优化进行代价聚合,提高匹配精度;采用赢家通吃法则进行视差计算,同时引入多步骤视差优化策略得到最终的视差图.在Middlebury测试平台提供的标准立体图像上,所提算法的平均误匹配率低至4.62％,与经典和新颖算法相比,匹配精度明显提高,具有更强的稳健性.
Abstract：Stereomatchingtechnologyisoneofthekeytechnologiesinthefieldofbinocularvision,anditisalsoafocusincurrentresearch.Itcalculatesthedisparityofmatchingpixelpairsinbinocularimagestoaccuratelyestimatethedepthofobjectsina3Dscene.However,inpractice,duetotheinfluenceofvariouscomplexfactorssuchaslensdistortion,lightingvariation,andocclusion,ithasbecomeachallengingtasktoaccuratelydeterminewhetherthecontentoftwoimagesmatches.Althoughthestereomatchingtechnologyhasmadesignificantprogressandarelativelycompletetechnicalsystemhasbeenestablished,therearestillsomeproblemsthatneedtobesolvedurgently.Amongthem,thetraditionalAD(AbsoluteDifference)-Censustransformationmethodisaprominentproblemcase.Thismethodisbasedonafixedmatchingwindowandreliestoomuchonthecenterpixelofthewindow,sowhenthecenterpixelisaffectedbyunfavorablefactorssuchaslighting,itsvalueispronetoabruptchanges,whichwillinterferewiththeaccuracyoftheAD-Censustransformationvalue.Thisinterferencenotonlyincreasestheriskofmismatchingthecorrespondingpixelsinthedisparitymap,butalsohasaseriousimpactonthestabilityofthealgorithm.Therefore,theexistingstereomatchingalgorithmsbasedonAD-Censushaveobviouslimitationsinobtainingthefinaldisparityresults,whichurgentlyrequiresthedevelopmentofanewalgorithmthatismorestableandrobust.Inordertoovercometheaboveproblems,thispaperproposesaninnovativematchingalgorithm,namelytheAD-Censusalgorithmcombiningmulti-pathgradientandreordering.Firstly,thetraditionalAD-CensustransformisrefinedbyreorderingpixelswithinvaryingscaletransformationwindowsandusingthemedianvalueasthecenterpixelforHammingdistancecalculations.Thismodificationimprovesrobustnesstonoiseandincorporateseight-directiongradientinformationforreliablecostcalculation.Furthermore,costaggregationisoptimizedthroughtheutilizationofcross-shapedsupportregionsandfour-pathscanlineoptimizations,elevatingmatchingprecision.Finally,theWinner-Take-All(WTA)strategyisadoptedfordisparitycalculation,complementedbyamulti-stepdisparityoptimizationapproachtoyieldtheultimatedisparityimage.ThenewalgorithminheritsthetraditionalAD-Censusmethodbasedonthewindowmatchingcharacteristicsandranksthepixelsinthematchingwindowbyweightingthem,andreplacesthecentralpixelsthatareabruptlyaffectedbynoisebytakingthemedianvalue.Indoingso,thematchingrelationshipbetweenpixelscanbemoreaccuratelyreflected,andtheinterferencecausedbytheabruptchangeofthecenterpixelintraditionalalgorithmscanbeeffectivelyreduced.Inaddition,theeight-directiongradientinformationfurtherimprovesthematchingaccuracyofthealgorithminthedepthdiscontinuousregionandweaktextureregionattheedgeoftheimage.Inordertoverifytheeffectivenessofthenewalgorithm,thispaperconductsextensiveexperimentalverificationonthestandardstereoimagepairontheMiddleburytestplatform.Theresultsshowthatcomparedwiththeclassicalandnovelstereomatchingalgorithms,theproposedalgorithmachievesasignificantreductioninthefalsematchingrate,reaching4.62％.Thisresultfullyprovestheexcellentperformanceofthenewalgorithmintermsofaccuracy.Atthesametime,thenewalgorithmalsoshowsobviousadvantagesintermsofrunningtime,whichfurtherprovesitshighefficiency.Ingeneral,theAD-Censusalgorithmcombiningmulti-pathgradientandreorderingproposedinthispapernotonlysuccessfullyovercomesthelimitationsoftraditionalalgorithms,butalsosignificantlyimprovesthematchingaccuracybyintroducinggradientinformation.Thisinnovativealgorithmhasinjectednewvitalityintothedevelopmentofstereomatchingtechnology,whichisexpectedtopromoteitswideapplicationinmanyfieldssuchas3Dreconstructionandrobotnavigation.
作者：林森 刘傲Author：LINSen LIUAo
作者单位：沈阳理工大学自动化与电气工程学院,沈阳110159
刊名：重庆理工大学学报 PKU
Journal：JournalofChongqingInstituteofTechnology
年，卷(期)：2024, 38(7)
分类号：TN201TP391.9
关键词：双目视觉 立体匹配 AD-Census 多路梯度
Keywords：binocularvision stereomatching ADCensusalgorithm eight-directiongradient
机标分类号：TP391.41TN911.73O212.6
在线出版日期：2024年6月24日
基金项目：国家重点研发计划，辽宁省教育厅高等学校基本科研项目结合多路梯度与重排序的AD-Census立体匹配算法[
期刊论文] 重庆理工大学学报--2024, 38(7)林森 刘傲针对局部立体匹配算法在深度不连续和弱纹理区域匹配精度低,且容易受到噪声干扰的问题,提出一种结合多路梯度与重排序的AD-Census立体匹配算法.对传统的AD-Census变换进行改进,重排序不同尺度变换窗口中的像素,取中值作...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
结合多路梯度与重排序的AD-Census立体匹配算法 AD-Census stereo matching algorithm combining multiple gradients and reordering
结合多路梯度与重排序的AD-Census立体匹配算法.pdf
- 文件大小:
- 4.28 MB
- 下载次数:
- 60
-
高速下载
|
|
