基于混合通道注意力的类别级物体六自由度位姿估计  

Category-level 6D object pose estimation based on mixed channel attention

作　　者：刘崇沛 孙炜[1] 刘剑[1] 杨慧 张星 范诗萌 Liu Chongpei;Sun Wei;Liu Jian;Yang Hui;Zhang Xing;Fan Shimeng(College of Electrical and Information Engineering,Hunan University,Changsha 410082,China)

机构地区：[1]湖南大学电气与信息工程学院,长沙410082

出　　处：《电子测量与仪器学报》2023年第7期72-80,共9页Journal of Electronic Measurement and Instrumentation

基　　金：国家自然科学基金(U22A2059);深圳科技计划项目(2021Szvup035);湖南大学汽车车身先进设计制造国家重点实验室自主研究项目;电子制造业智能机器人技术湖南省重点实验室开放课题项目资助。

摘　　要：针对有光照变化、距离变化、背景杂乱、遮挡等干扰的场景下物体六自由度位姿估计精度低的问题,提出了一种结合多尺度特征融合和注意力机制的混合通道注意力模块(mixed channel attention,MCA)。在MCA基础上进一步构建了类别级物体六自由度位姿估计方法(MCA6D),其关键步骤包括物体的实例分割,特征提取与基于MCA的特征优化,基于先验形状的物体模型重建,及基于点云配准的位姿估计。本文方法在公共数据集CAMERA和REAL分别取得86.3%(5°2 cm)、73.4%(5°5 cm)和39.2%(5°2 cm)、43.3%(5°5 cm)的均值平均精度,领先于NOCS、SPD、SGPA等主流方法;同时实物实验表明本文方法在存在光照变化、距离变化、背景杂乱、遮挡等干扰的场景可准确估计物体六自由度位姿。Aiming at the low accuracy of object six-degree-of-freedom(6D)pose estimation in scenes with interferences such as illumination changes,distance changes,background clutter,and occlusions,a mixed channel attention module(MCA)is proposed,which combines multi-scale feature fusion and attention mechanisms.Based on MCA,a category-level object 6D pose estimation method(MCA6D)is further constructed.The key steps include object instance segmentation,feature extraction and optimization based on MCA,object model reconstruction based on prior shape,and pose estimation based on point cloud registration.Relevant experiments show that our method achieves 86.3%(5°2 cm),73.4%(5°5 cm)and 39.2%(5°2 cm),43.3%(5°5 cm)mean average precision in the public datasets CAMERA and REAL,respectively,which is ahead of mainstream methods such as NOCS,SPD,and SGPA.At the same time,the practical experiment shows that the proposed method can accurately estimate the 6D pose of the object in scenes with interference,such as illumination changes,distance changes,background clutter,and occlusions.

关 键 词：物体六自由度位姿估计 类别级 注意力机制 通道注意力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN0[自动化与计算机技术—计算机科学与技术]