基于稠密向量图的物体6D位姿回归算法  

6D Object Pose Regression Method Based on Dense Vector Graphs

作　　者：左国玉 喻杉 顾宗函 郑榜贵 ZUO Guoyu;YU Shan;GU Zonghan;ZHENG Banggui(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computional Intelligence and Intelligent System,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]计算智能与智能系统北京市重点实验室,北京100124

出　　处：《北京工业大学学报》2024年第12期1409-1420,共12页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(62373016);多模态人工智能系统全国重点实验室开放课题资助项目(MAIS-2023-22)。

摘　　要：针对稀疏间接两阶段位姿估计算法遮挡、噪声鲁棒性差、实时性低以及n点透视投影(perspective-n-point,PnP)计算过程不可微分等难题,提出基于稠密向量图的端到端物体6D位姿回归算法。以投票算法第1阶段预测的稠密张量信息为输入,代替逐像素投票及PnP计算环节,端到端直接回归物体6D位姿。首先,在特征提取模块提取成对向量特征;其次,在特征聚合模块引入基于余弦相似函数的聚类特征损失正则项;最后,通过旋转、平移解耦推理物体6D位姿。在网络训练的数据输入环节加入随机噪声数据以提升算法的噪声鲁棒性,在球体合成数据集和Occlusion-LINEMOD数据集上的测试结果证明了该算法的有效性。To solve the problem that sparse two-stage object pose estimation methods have poor occlusion robustness and real-time performance,as well as poor noise robustness and non differentiable perspectiven-point(PnP)calculation process,an end-to-end 6D object pose regression method based on dense vector fields was proposed.The method regards the dense tensor information predicted in the first stage of the voting-based method as input and replaces the pixel-wise voting and PnP calculation processes of the voting-based method to end-to-end regress 6D object pose information.First,paired vector features were extracted in the feature extraction module.Then,clustering feature loss based on cosine similarity functions was introduced in the feature aggregation module.Finally,the 6D object pose was estimated through rotation and estimation decoupling.Random noise data was added during the data input stage of network training process to improve the noise robustness of the method.The results on the sphere synthesis dataset and the public dataset Occlusion-LINEMOD prove the validity of this method.

关 键 词：深度学习 位姿估计 特征聚类 位姿解耦 稠密向量图 端到端 

分 类 号：U621.3[交通运输工程—船舶及航道工程] TP183[交通运输工程—船舶与海洋工程]