零件点云法向量估计的多尺度特征融合网络

为了解决机械零配件点云处理中非均匀采样干扰、尖锐特征损失等难点,提出一种基于深度神经网络多尺度融合的点云法向量估计方法.该网络在不同邻域尺度下集成了采样点细节与点云块整体两种特征.为了使该多维回归输出网络的训练更稳定且能缓解梯度爆炸问题,重新设计了一个光滑的损失函数.实验结果表明,该方法性能优于传统的方法以及HoughCNN、PCPNet等方法,能够更准确地估计尖锐边缘的法向量,对点云各种噪声和采样方法鲁棒性都更强.

Multi-scale feature fusion network for point cloud normal estimation of mechanical parts

Aimed to handle the challenging difficulties of sampling irregularity and sharp features lost when processing point clouds of mechanical parts, this paper proposed a multi-scale feature fusion strategy of normal estimation based on deep learning neural network. The network aggregated local point features and point cloud patch properties in different scales. Besides, this paper introduced a novel loss function in order to improve the stability and prevent exploding gradients in 3D regression network training. Experimental results indicate that this method is consistently better than PCA, HoughCNN and PCPNet. This model can accurately estimate the normal in sharp edges, while remaining robust to noise and sampling density.