首页|基于改进注意力机制和曲面微分几何的2阶段点云重建

基于改进注意力机制和曲面微分几何的2阶段点云重建

Two-stage Point Cloud Reconstruction Based on Improved Attention Mechanism and Surface Differential Geometry

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
针对单个阶段直接预测点云的直通式结构会导致生成的点云相对稀疏、细节描述不清晰、结果分布不均匀等问题,提出一种基于改进注意力机制和局部曲面微分几何的 2 阶段密集点云重建网络,实现分阶段预测不同分辨率、分布均匀的密集点云.首先通过嵌入改进坐标注意力机制,提高网络对输入图像中目标的方向感知和坐标信息的捕获能力;然后预测出目标的稀疏点云,并提取稀疏点云的密集连接特征获得描述稀疏点云的结构感知信息;最后采用局部曲面微分几何完成点云扩展进而生成分布均匀的密集点云.实验结果表明,与 3D-FEGNet相比,所提密集点云重建网络在CD和EMD上分别降低 25.4%和 25.1%,采用真实物体进行实验,所获得点云三维尺寸误差均<1.5 mm.
The straight-through structure of direct prediction of point clouds in a single stage will lead to relatively sparse point clouds,unclear details and uneven distribution of results.Therefore,a 2-stage dense point cloud reconstruction network based on improved attention mechanism and local surface differential geometry is proposed,which can realize phased prediction of dense point clouds with different resolutions and uniform distribution.Firstly,the coordinate attention mechanism is improved by embedding to improve the network's ability to perceive the direction of the target and capture the coordinate information in the input image.Then,the sparse point cloud of the target is predicted,and the dense connection features of the sparse point cloud are extracted to obtain the structural perception information describing the sparse point cloud.Finally,the local surface differential geometry is used to complete the point cloud expansion,and then generate a dense point cloud with uniform distribution.The experimental results show that compared with 3D-FEGNet,the proposed dense point cloud reconstruction network reduces by 25.4%and 25.1%on CD and EMD,respectively.The three-dimensional dimensional errors of point clouds obtained by experiments with real objects are all less than 1.5 mm.

machine visionpoint cloud reconstructiondeep learningsingle imagelocal surface differential geometrydense point cloudsreconstruction network

王凯、陈辉、陈廉明、黄和平、陈晓琳

展开 >

上海电力大学 自动化工程学院,上海 200090

浙江正泰仪器仪表有限责任公司杭州分公司,浙江 杭州 310052

上海正泰电源系统有限公司,上海 201614

机器视觉 点云重建 深度学习 单幅图像 局部曲面微分几何 密集点云 重建网络

国家科技部外国专家局项目上海市科委科技创新双碳项目

DL2022013007L21DZ1207300

2024

10.3969/j.issn.1000-1158.2024.07.04

计量学报
中国计量测试学会

计量学报

CSTPCD北大核心
影响因子：0.303
ISSN：1000-1158
年,卷(期)：2024.45(7)
  • 5