多层特征融合与混合注意力的物体位姿估计

Object Pose Estimation Based on Multi-Layer Feature Fusion and Hybrid Attention

在工业机器人抓取过程中,针对物体无纹理、存在遮挡、场景杂乱的情况下的6D位姿精确估计问题,提出一种多层特征融合与混合空间通道注意力的物体6D位姿估计算法。设计了一种垂直连接的双向特征融合金字塔网络,实现多层特征融合,提升对目标关键点的检测性能。嵌入混合空间通道注意力机制,聚焦空间和通道两个维度上的特征信息,增强模型的局部表征能力。在LineMod数据集及Occlusion LineMod遮挡数据集上的实验结果表明所提出算法的优越性及有效性,且能够有效处理背景杂乱及遮挡问题。

In the process of industrial robot grasping,aiming at the problem of 6D pose estimation accuracy under the condition of no texture,occlusion and chaotic scene,a 6D pose estimation algorithm for objects that mixes spatial channel attention and multi-layer feature fusion is proposed.A vertically connected bidirectional feature fusion pyramid network is designed to realize multi-layer feature fusion and improve the detection performance of target key points.Embedded in the hybrid spatial channel attention mechanism,focusing on the feature information in the two dimensions of space and channel,and enhancing the local representation ability of the model.Experimental results on LineMod dataset and Occlusion LineMod occlusion dataset show the superiority of the proposed algorithm,and can effectively deal with background clutter and occlusion problems.