基于方向编码与空洞采样的室内点云物体分割

Indoor Point Cloud Object Segmentation Based on Direction Coding and Hole Sampling

针对现有方法在处理局部特征时忽略方向信息,且由于卷积核大小的限制无法有效地提取点云邻域特征等问题,提出一种点云分割方法.首先结合方向编码和空洞采样最大程度扩大网络的局部感受野.其次利用图卷积神经网络挖掘局部邻域内点的信息.然后使用邻域特征提取层自动加权融合邻域特征为更具有代表性的单个特征点.最后结合空间注意力机制,增加远程点之间的联系.在S3DIS数据集上进行物体分割实验的结果表明,所提方法的OA和mIoU比PointWeb高1.3个百分点和4.0个百分点,比基线方法 RandLA-Net高0.6个百分点和0.7个百分点,使用空洞采样与方向编码能够有效地提高点云的语义分割精度.

To address the problem that existing methods ignore directional information when processing local features and cannot effectively extract point cloud neighborhood features due to the limitation of convolutional kernel size,this paper proposes a point cloud segmentation method.Firstly,by combining directional encoding and hole convolution,the local receptive field of the network is maximized.secondly,graph convolutional neural networks are utilized to mine information within local neighborhoods of points.subsequently,a neighborhood feature extraction layer is employed to automatically weight and fuse neighborhood features into a more representative single feature point.finally,in conjunction with a spatial attention mechanism,the connections between distant points are enhanced.The results of object segmentation experiments on the S3DIS dataset show that the OA and mIoU of the proposed method are,respectively,1.3 percentage points and 4.0 percentage points higher than PointWeb,0.6 percentage points and 0.7 percentage points higher than the baseline method RandLA-Net,and the use of hole sampling and directional coding can effectively improve the semantic segmentation accuracy of point clouds.