相对位姿与姿态池优化的类级别6D姿态估计模型

打开文本图片集

中图分类号：TP391 文献标志码：ADOI：10.12475/aie.20250602

本文引用格式：，，，等.相对位姿与姿态池优化的类级别6D姿态估计模型[J].自动化与信息工程，2025，46（6）：9-16.HUANG Tianlun，HUShengchao，LIUXiao，etal.Relative poseand pose pool optimizedcategory-level6Dposeestimation model[J].Automation& Information Engineering，2025，46（6）：9-16.

文章编号：1674-2605（2025）06-0002-08开放获取

Relative Pose and Pose Pool Optimized Category-level 6DPoseEstimation Model

HUANG Tianlun1，2 HU Shengchao1，2 LIU Xiaol WANG Weijun1，2 FENG Wei2 （1.Shenzhen Institutes ofAdvanced Technology，ChineseAcademyof Sciences，Shenzhen 518055，China 2.University of Chinese Academy of Sciences，Beijing 101408， Chia）

Abstract：Toaddressthe chalenges faced by6Dpose estimation incomplex environments—suchas heavy reliance on highprecision3Dodelsorlage-saletemplatelibaries，severeoclusion，andlightinginterferecetispaperpropossaategor-level 6Dposeestimationmodelbasedonrelativeposeandposepooloptimization.Themodelconstructsaprogressiveestimationframework encompassingettgtadeiottlatieellatiodeolsdio Bycomputing theabsoluteposeofthetargetobjectthroughrelativeposesbetweenadjacentframes，itreducesdependencyonhighprecision3DodelsortmpatelrariesFurthmoreoseoolotiationratgyistroduceditegatigself-ion cross-atentionmechanisms tofusehistoricalframeinformation，therebyenhancingtheaccuracyndobustnessofthecuetpose estimation.Experimentalresults on public datasets NOCS and HO3Ddemonstrate thatthe modelachieves anaverage 5∘5 cm accuracy of 90.35% ，anaverage IoU25of 99.97% ，and an average translation error of 1.72cm on the NOCS dataset，outperforming three comparative models.The average rotation error is 1.90∘ ，which is comparable to the BundleSDF model. On the HO3D dataset， the model achieves an ADD-S of 96.9% and an ADD of 93.4% ，surpassing four comparative models.These results indicate that the model canachieve accurateandstablecategory-level 6Dpose estimatioevenunder weakmodeldependency，providingcrucialsupportfor robotic manipulation in complex environments and practical deployment.

Keywords： 6Dpose estimation;relative pose; pose-pooloptimization; category-levelpose estimation;atention mechanism

0 引言

态估计提供关键支撑。（剩余9497字）