In order to address the pose estimation challenges encountered in robotic grasping due to object symmetry and occlusions, the problem of six-degree-of-freedom pose estimation was investigated by integrating multi-view RGB-D data. The proposed method was built upon a single view pose estimation framework, in which a 3D keypoints voting network was employed to independently predict 3D keypoints from each viewpoint. Multi-view information was then effectively integrated through confidence scoring, weighted fusion, and manshift clustering, enabling accurate recovery of the target object pose. Experimental results demonstrate that the proposed method significantly improves pose estimation accuracy in symmetric and occluding scenarios. The multi-view fusion strategy effectively enhances stability and robustness of keypoints prediction. Autonomous grasping experiments in tabletop environments achieve notably higher grasp success rates and execution performance. The experiments have shown that the proposed method has good adaptability and application value in complex grasping scenarios, can effectively improve the autonomy of operations in unstructured scenarios, and provides key technical support for solving intelligent assembly operations under complex working conditions, with broad prospects for industrial application.
YANG Hong-ling, DAI Lu, XIE Lun.
Robot Grasping Pose Estimation Model Based on Multi-view Fusion[J].
Science Technology and Engineering, 2026, 26(9): 3616-3625 DOI:10.12404/j.issn.1671-1815.2506992
目标物体的6DoF位姿估计方法通常需要结合目标检测或实例分割,提取出目标物体的感兴趣区域(region of interest, ROIs),再对ROIs内的目标物体进行位姿估计,PVN3D同时完成多类物体的多个实例对象的分割与位姿估计任务。为达此目的,该方法对实例分割模块Ms、中心点预测模块Mc及关键点预测模块Mk进行联合优化,提升模型性能。实例分割模块Ms引导网络同时获取全局与局部特征,这些特征能够辅助关键点预测模块Mk确定的平移信息,帮助实例分割模块区分外同一类别的目标物体;同时中心点预测模块Mc辅助Ms实现对不同实例的区分。PVN3D利用监督学习方法,对3个模块Mk、Ms以及Mc进行了协调优化,其最终的损失函数为
以与PVN3D一致的测试集展开验证工作,该测试集中涵盖了YCB-Video数据集中的所有21个物体,共计2 949帧。在该测试集中,重点聚焦于存在遮挡以及对称情况的物体,具体包括tomato soup can (005)、potted meat can (010)、bowl (024)、wood block (036)、large clamp (051) 以及 extra large clamp (052)。与PVN3D相同,采用平均距离偏差(average distance of model points, ADD)和对称物体平均距离偏差(average closest point distances, ADD-S)作为目标物体6DoF位姿估计误差评估标准。物体模型的点云数据经过模型预测出的位姿(R, t)变换转换为预测点云,真实位姿经过(R*, t*)转换为真实点云。ADD对模型预测位姿的误差进行评估,其方式是算出预测点云与真实点云对应点之间的平均欧式距离,表达式为
为了验证位姿估计方法的表现,采用了与PVN3D相一致的评估方式,采用ADD(S) AUC (the area under the accuracy-threshold curve, AUC)和ADD-S AUC两种评估方式,二者的区别在于ADD(S) AUC对非对称物体计算ADD距离和对称物体计算ADD-S距离,而ADD-S AUC则统一计算ADD-S距离。
从量化指标来看,基于多视角融合的位姿估计算法相比原始PVN3D模型在对称及遮挡物体(用*表示)tomato soup can (005)、potted meat can (010)、bowl (024)、wood block (036)、large clamp (051)以及extra large camp (052)上平均提升5.2%,最高提升8.8%,最低提升2.4%。在整个测试集的平均性能相比原始PVN3D在ADD(S) AUC和ADDS AUC指标提高2.4%和1.8%。
3.2 场景关键点置信度评分
场景关键点置信度评分过程如图5所示,假定图5(a)和图5(b)分别为视角k下的RGB图像及场景点云,同时物体Master Chef Can记为物体i,如图5(a)中红色矩形框所示。令物体i上第j个场景关键点为$k{p}_{j}^{k,i}$,图5(c)中的绿色点云为实例分割结果。其属于第i个物体的场景点云投票出的场景关键点$k{p}_{j}^{k,i}$的空间位姿,绿色点云的数目即为nk,i。然后,对场景关键点$k{p}_{j}^{k,i}$的投票结果进行聚类,如图5(d)所示,聚类中心即为单视角下场景关键点$k{p}_{j}^{k,i}$的最终位姿,图中红色点云即为聚类结果中处于以场景关键点$k{p}_{j}^{k,i}$为中心以r为半径的球体内的点,本文研究中r=0.02。图5(d)中红色点云数目与绿色点云数目的比值即为场景关键点$k{p}_{j}^{k,i}$的置信度评分。
LeiZhiming, LiYong, ShuangFeng, et al. Pose estimation algorithm for robot arm target grabbing based on RGB image[J]. Computer Integrated Manufacturing Systems, 2023, 29(10): 3296.
ZhangRihong, ChenDezhao, ZhangRuihua, et al. Review of research on cooperative picking technology with multiple robotic arms[J]. Science Technology Engineering, 2025, 25(15): 6155-6168.
[6]
HolenH KR, SmithT, EgelandO. Visual estimation of relative ship-crane movement from point and line correspondences[J]. Journal of Offshore Mechanics Arctic Engineering, 2026, 148(2): 021402.
[7]
HeT, FangY, LiK, et al. Large-scale point cloud completion through registration and fusion of object-level reconstructions[J]. Applied Sciences, 2026, 16(1): 554.
[8]
JovhariN, SedaghatA, Shah-HosseiniR, et al. Robust retinal image matching: a modality-resistant descriptor using directional anisotropic texton-like features and evaluation across diverse datasets[J]. Journal of Imaging Informatics in Medicine, 2026(1): 1-24.
[9]
YuZ, WangH, WuY, et al. A convenient height-filtering-based strategy for multi-source forest lidar point cloud registration[J]. Smart Agricultural Technology, 2026, 13: 101790.
[10]
HeY, WangS, WuC, et al. PGA6D: 6D pose estimation for grasping and assemblying based on keypoints voting[J]. Machine Vision Applications, 2026, 37(1): 11.
[11]
RanH, ChenD, ChenQ, et al. 6-DoF grasp estimation method that fuses RGB-D data based on external attention[J]. Journal of Visual Communication Image Representation, 2024, 101: 104173.
[12]
De RoovereP, DaemsR, CroenenJ, et al. CCPose: high-precision six-dimensional pose estimation for industrial objects[J]. Machine Vision Applications, 2026, 37(1): 2.
[13]
PengS, LiuY, HuangQ, et al. PVNet: pixel-wise voting network for 6DoF pose estimation[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 4561-4570.
[14]
WangC, XuD, ZhuY, et al. Densefusion: 6D object pose estimation by iterative dense fusion[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 3343-3352.
[15]
ChenW, JiaX, ChangH, et al.G2l-net: global to local network for real-time 6D pose estimation with embedding vector features[C]. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 4233-4242.
[16]
HeY, SunW, HuangH, et al.PVN3D: a deep point-wise 3D keypoints voting network for 6DoF pose estimation[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 11632-11641.
[17]
LinH, LiangS, ChenY. Robotic grasping with multi-view image acquisition and model-based pose estimation[J]. IEEE Sensors Journal, 2020, 21(10): 11870-11878.
[18]
LiuY, ZhangCDongX, et al. Point cloud-based deep learning in industrial production: a survey[J]. ACM Computing Surveys, 2025, 57(7): 1-36.
[19]
ZhuZ, PengS, LarssonV, et al. NICER-SLAM: neural implicit scene encoding for RGB SLAM[C]. International Conference on 3D Vision (3DV). Piscataway: IEEE, 2024: 42-52.
[20]
PlacedJ A, StraderJ, CarrilloH, et al. A survey on active simultaneous localization and mapping: state of the art and new frontiers[J]. IEEE Transactions on Robotics, 2023, 39(3): 1686-1705.
[21]
LuoF, ZhuY, FuY, et al. Sparse RGB-D images create a real thing: a flexible voxel based 3D reconstruction pipeline for single object[J]. Visual Informatics, 2023, 7(1): 66-76.
[22]
ZhangX, JiangX, HanJ, et al. Path planning method for stacking parts of industrial robots guided by 3D vision[J]. Journal of Computers, 2024, 35(4): 225-240.
[23]
WangC, PeiZ, QiuS, et al. RGB-D-based stair detection and estimation using deep learning[J]. Sensors, 2023, 23(4): 2175.
[24]
TangT, HoC C, NadarajanG. Markerless feature-registration-based image processing framework for oil palm fresh fruit bunch robotic harvester[C]// International Workshop on Advanced Imaging Technology (IWAIT). Bellingham: SPIE, 2025: 127-132.
[25]
HuangL, BaiJ, GuoJ, et al. On the error analysis of 3d gaussian splatting and an optimal projection strategy[C]// European conference on computer vision (ECCV). Cham: Springer, 2024: 247-263.
[26]
YangH, WangY, YinX, et al. A 6-dimensional pose estimation method combining sparse viewpoint classification initialization and optical flow-guided iterative refinement[J]. Engineering Applications of Artificial Intelligence, 2025, 159: 111814.
[27]
LuoJ, CaiJ, LiJ, et al. Detection and pose measurement of underground drill pipes based on GA-PointNet++[J]. Applied Intelligence, 2025, 55(2): 132.
[28]
RajičF, XuH, MihajlovicM, et al. Multi-view 3D point tracking[C]// IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2025: 59-68.
[29]
LinY, TremblayJ, TyreeS, et al. Multi-view fusion for multi-level robotic scene understanding[C]// IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Piscataway: IEEE, 2021: 6817-6824.