Pose estimation based on RGB-D images is a hot issue that has attracted much attention in recent years. A key technical challenge is to extract features from depth information and image information separately and fully leverage the two complementary data sources. The previous methods ignored the internal connection of local features and the feature fusion of heterogeneous data, limiting the robustness and real-time performance in cluttered scenes. In this article, we propose LHFF-Net, a generic framework based on dynamic graph convolution to strengthen the information aggregation among all point clouds in a local region. After extracting heterogeneous features, we fuse information from two data sources in different receptive fields, to estimate the pose of the object while fully extracting local features. We show in experiments that the proposed approach outperforms state-of-the-art approaches on two challenging data sets, YCB-Video and LineMOD. We also have deployed our proposed method on the UR5 robot for grasping experiments and achieved good grasping performance.

基于RGB-D图像的姿态估计是近年来备受关注的热点问题。一个关键的技术挑战是分别从深度信息和图像信息中提取特征，并充分利用这两个互补的数据源。以往的方法忽略了局部特征的内在联系和异构数据的特征融合，限制了在杂乱场景中的鲁棒性和实时性。在本文中，我们提出了 LHFF-Net，这是一种基于动态图卷积的通用框架，用于加强局部区域内所有点云之间的信息聚合。在提取异构特征后，我们融合来自不同感受野的两个数据源的信息，在充分提取局部特征的同时估计对象的姿态。我们在实验中表明，所提出的方法在两个具有挑战性的数据集 YCB-Video 和 LineMOD 上优于最先进的方法。我们还在 UR5 机器人上部署了我们提出的方法进行抓取实验，并取得了良好的抓取性能。