基于深度學(xué)習(xí)的姿態(tài)估計(jì)研究的國(guó)內(nèi)外文獻(xiàn)綜述姿態(tài)估計(jì)的目的是對(duì)工作環(huán)境中目標(biāo)物體的位置和朝向進(jìn)行計(jì)算，為機(jī)器人的抓取操作、三維重建[[]SalisburyJR,DeverellMH,CooksonMJ.Three-dimensionalreconstruction[J].ElectronicJournalofPathology&Histology,1993,8(4):24041-240045.]、人機(jī)交互等操作提供基礎(chǔ)信息，在增強(qiáng)現(xiàn)實(shí)[[][]SalisburyJR,DeverellMH,CooksonMJ.Three-dimensionalreconstruction[J].ElectronicJournalofPathology&Histology,1993,8(4):24041-240045.[]ChuiYP,HengPA.EnhancingViewConsistencyinCollaborativeMedicalVisualizationSystemsUsingPredictive-BasedAttitudeEstimation[C]//MedicalImagingandAugmentedReality:FirstInternationalWorkshop,MIAR2001,HongKong,China,June10-12,2001.Proceedings.IEEE,2001.1.1基于模板匹配的位姿估計(jì)方法研究現(xiàn)狀基于模板匹配的算法是通過(guò)對(duì)現(xiàn)實(shí)環(huán)境中的目標(biāo)物體的特征和模型庫(kù)中現(xiàn)有模型進(jìn)行特征匹配，得到目標(biāo)物體的六自由度位姿估計(jì)。Hinterstoisser在2011年提出的模板匹配姿態(tài)估計(jì)方法Linemod算法[]HinterstoisserS,HolzerS,CagniartC,etal.Multimodaltemplatesforreal-timedetectionoftexture-lessobjectsinheavilyclutteredscenes[C]//IEEEInternationalConferenceonComputerVision.IEEE,2011:858-865.，利用目標(biāo)物體的顏色梯度信息和深度圖像中物體表面的法向量特征[][]HinterstoisserS,HolzerS,CagniartC,etal.Multimodaltemplatesforreal-timedetectionoftexture-lessobjectsinheavilyclutteredscenes[C]//IEEEInternationalConferenceonComputerVision.IEEE,2011:858-865.[]SunW,WangJ,JinF,etal.DATUMFEATUREEXTRACTIONANDDEFORMATIONANALYSISMETHODBASEDONNORMALVECTOROFPOINTCLOUD[C]//2018.[]蔡靜,韓丹,張琰,等.基于分層分割的滑動(dòng)窗搜索方法:,CN103870834B[P].2017.[]

VetterT,RomdhaniS,PierrardJS.Estimating3Dshapeandtextureofa3Dobjectbasedona2Dimageofthe3Dobject:US,US7756325B2[P].2010.[]HodaňT,ZabulisX,LourakisM,etal.Detectionandfine3Dposeestimationoftexture-lessobjectsinRGB-Dimages[C]//2015IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems(IROS).IEEE,2015:4421-4428.圖1.2LINEMOD表面特征示意總之，基于模板匹配的姿態(tài)估計(jì)算法是通過(guò)對(duì)比真實(shí)物體的幾何特征模型圖像與觀測(cè)圖像的特征，對(duì)比RGB圖象和模板RGB圖象的梯度信息，根據(jù)匹配最相似模板圖像來(lái)估計(jì)觀測(cè)目標(biāo)物體的姿態(tài)。1.2基于投票的目標(biāo)姿態(tài)估計(jì)方法研究現(xiàn)狀基于投票的目標(biāo)姿態(tài)估計(jì)算法主要應(yīng)用于目標(biāo)物體存在遮擋情況時(shí)的位姿估計(jì)問(wèn)題，Drost等人[]DrostB,IlicS.3DObjectDetectionandLocalizationUsingMultimodalPointPairFeatures.IEEEComputerSociety,2012.提出了一種PointPairFeatures算法，算法原理如圖1.3所示，點(diǎn)特征包含三維點(diǎn)之間的位姿信息，算法中沒(méi)有采用三維特征描述子這樣的局部特征，而是采用了一種全局的點(diǎn)對(duì)特征[]魯榮榮,朱楓,吳清瀟,等.基于增強(qiáng)型點(diǎn)對(duì)特征的三維目標(biāo)識(shí)別方法[J].光學(xué)學(xué)報(bào),2019,v.39;No.449(08):245-254.。Kim等人[]KimJ,JunH.ReducingFeaturePointMismatchesusingMultipleDescriptors.在初始的點(diǎn)對(duì)特征中融入了空間、表面與不可見(jiàn)表面等可見(jiàn)特征，增強(qiáng)了算法的點(diǎn)對(duì)特征匹配能力。Branchmann等人[]BrachmannE,KrullA,MichelF,etal.Learning6DObjectPoseEstimationUsing3DObjectCoordinates[M].SpringerInternationalPublishing,2014.提出了ObjectCoordinates算法，該算法提出了一個(gè)新的概念，使用RandomForest[]DrostB,IlicS.3DObjectDetectionandLocalizationUsingMultimodalPointPairFeatures.IEEEComputerSociety,2012.[]魯榮榮,朱楓,吳清瀟,等.基于增強(qiáng)型點(diǎn)對(duì)特征的三維目標(biāo)識(shí)別方法[J].光學(xué)學(xué)報(bào),2019,v.39;No.449(08):245-254.[]KimJ,JunH.ReducingFeaturePointMismatchesusingMultipleDescriptors.[]BrachmannE,KrullA,MichelF,etal.Learning6DObjectPoseEstimationUsing3DObjectCoordinates[M].SpringerInternationalPublishing,2014.[]LiawA,WienerM.ClassificationandRegressionbyrandomForest[J].RNews,2002,23(23).[]Visuallytrackinganobjectinrealworldusing2dappearanceandmulticuedepthestimations[J].US,2014.圖1.2PPF算法原理1.3基于深度學(xué)習(xí)的目標(biāo)姿態(tài)估計(jì)方法研究現(xiàn)狀隨著神經(jīng)網(wǎng)絡(luò)和深度學(xué)習(xí)技術(shù)的發(fā)展，由于卷積神經(jīng)網(wǎng)絡(luò)強(qiáng)大的特征提取能力，越來(lái)越多的人開(kāi)始對(duì)端到端的深度模型結(jié)合6D姿態(tài)估計(jì)進(jìn)行研究，越來(lái)越多基于神經(jīng)網(wǎng)絡(luò)的物體位姿估計(jì)方法也不斷發(fā)展。YXiang等人于2017年提出一種新穎的PoseCNN姿態(tài)估計(jì)算法，使用卷積神經(jīng)網(wǎng)絡(luò)提取輸入圖像中分辨率為原來(lái)分辨率1/8和1/16的兩個(gè)特征圖，然后在網(wǎng)絡(luò)中將特征圖通過(guò)反卷積操作將其恢復(fù)為輸入圖像大小的特征圖，并生成像素的語(yǔ)義標(biāo)記[]MartinSC.Mappingofsemantictagstophasesforgrammargeneration[J].US,2008.，估計(jì)3D位移時(shí)，為了避免目標(biāo)遮擋帶來(lái)的目標(biāo)中心檢測(cè)失效，在ImplicitShapeModel[[]MartinSC.Mappingofsemantictagstophasesforgrammargeneration[J].US,2008.[]LeibeB.Combinedobjectcategorizationandsegmentationwithanimplicitshapemodel[C]//Proc.ECCVworkshoponstatisticallearningincomputervision,2004.2004.圖1.3PoseCNN工作流程Rad等人[]RadM,LepetitV.BB8:AScalable,Accurate,RobusttoPartialOcclusionMethodfor\n,Predictingthe3DPosesofChallengingObjectswithoutUsingDepth[C]//2017IEEEInternationalConferenceonComputerVision(ICCV).IEEE,2017.在2017年提出了一種僅使用RGB圖像的姿態(tài)估計(jì)算法BB8，該算法使用RGB圖像作為輸入，該算法先以分割[]韓思奇,王蕾.圖像分割的閾值法綜述[J].系統(tǒng)工程與電子技術(shù),2002,24(6):91-94.的方式檢測(cè)環(huán)境中的目標(biāo)物體的2D圖像，再利用目標(biāo)物體的2D投影預(yù)測(cè)姿態(tài)。這種方式使得其在處理有旋轉(zhuǎn)對(duì)稱(chēng)軸的物體時(shí)效果不佳，為了彌補(bǔ)這一缺點(diǎn)，該算法引入了分類(lèi)器[]BottouL,CortesC,DenkerJS,etal.Comparisonofclassifiermethods:acasestudyinhandwrittendigitrecognition[C]//InternationalConferenceonPatternRecognition.IEEEComputerSociety,1994.來(lái)解決問(wèn)題。Kehl等人[]RadM,LepetitV.BB8:AScalable,Accurate,RobusttoPartialOcclusionMethodfor\n,Predictingthe3DPosesofChallengingObjectswithoutUsingDepth[C]//2017IEEEInternationalConferenceonComputerVision(ICCV).IEEE,2017.[]韓思奇,王蕾.圖像分割的閾值法綜述[J].系統(tǒng)工程與電子技術(shù),2002,24(6):91-94.[]BottouL,CortesC,DenkerJS,etal.Comparisonofclassifiermethods:acasestudyinhandwrittendigitrecognition[C]//InternationalConferenceonPatternRecognition.IEEEComputerSociety,1994.[]KehlW,ManhardtF,TombariF,etal.SSD-6D:MakingRGB-Based3DDetectionand6DPoseEstimationGreatAgain[C]//2017IEEEInternationalConferenceonComputerVision(ICCV).IEEE,2017.[]LiuW,AnguelovD,ErhanD,etal.SSD:SingleShotMultiBoxDetector[J].Springer,Cham,2016.Wang等人[]WangC,DXu,ZhuY,etal.DenseFusion:6DObjectPoseEstimationbyIterativeDenseFusion[J].2019.提出使用了一種新的密集融合網(wǎng)絡(luò)來(lái)進(jìn)行姿態(tài)估計(jì)的方法，首先進(jìn)行語(yǔ)義分割，得到點(diǎn)云信息和被mask裁剪的圖像，然后使用全卷積網(wǎng)絡(luò)提取圖形的顏色特征，并且用PointNet[[]WangC,D