| 摘 要: 针对复杂点云中多尺度几何信息表达不足、局部结构关系建模能力有限以及跨点云特征一致性难以建立等问题，研究了一种基于PointNet改进的复杂点云特征提取网络。该方法在PointNet基础上引入多尺度特征提取模块、局部结构增强模块和跨点云注意力机制匹配模块，以增强网络对多尺度上下文信息、局部几何结构和跨云关联特征的表达能力，并采用交叉熵损失函数进行监督训练。以ModelNet40数据集和斯坦福三维扫描库为实验对象，通过对应点平均特征距离、非对应点平均特征距离、最近邻匹配准确率和匹配召回率等指标进行评价。实验结果表明，所提方法在复杂点云特征提取与跨点云匹配任务中均优于原始PointNet网络，验证了该方法具有较好的特征表达能力和匹配稳定性。 |
| 关键词: 深度学习 3D点云 特征提取 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金民航联合基金重点项目(U243320053) |
|
| Research on a complex point cloud feature extraction network improved based on PointNet |
|
yeyuxiang, fuyaoming
|
College of Aviation Engineering， Civil Aviation Flight University of China
|
| Abstract: To address the insufficient representation of multi-scale geometric information, limited modeling capability of local structural relationships, and difficulty in establishing feature consistency across complex point clouds, this paper investigates an improved PointNet-based feature extraction network for complex point clouds. On the basis of PointNet, the proposed method introduces a multi-scale feature extraction module, a local structure enhancement module, and a cross-point-cloud attention matching module to enhance the representation of multi-scale contextual information, local geometric structures, and cross-cloud feature associations. A cross-entropy loss function is adopted for supervised training. Experiments are conducted on the ModelNet40 dataset and the Stanford 3D Scanning Repository. The proposed method is evaluated using the average feature distance of corresponding points, the average feature distance of non-corresponding points, nearest-neighbor matching accuracy, and matching recall. The experimental results show that the proposed method outperforms the original PointNet in complex point cloud feature extraction and cross-point-cloud matching tasks, demonstrating better feature representation capability and matching stability. |
| Keywords: Deep learning 3D point cloud feature extraction |
