基于区块特征融合的点云语义分割方法

为降低室外大规模点云场景中多类三维目标语义分割的计算复杂度，提出一种融合区块特征的语义分割方法。采用方形网格分割方法对三维点云进行区块划分、采样以及组合，求取简化的点云组合区块集，将其输入至区块特征提取和融合网络中从而获得每个区块的特征修正向量。设计点云区块全局特征修正网络，以残差的方式融合特征修正向量与原始点云全局特征，修正因分割造成的错误特征。在此基础上，将方形网格分割尺寸作为神经网络的参数引入反向传播过程中进行优化，从而建立高效的点云语义分割网络。实验结果表明，反向传播算法可以优化分割尺寸至最佳值附近，所提网络中的全局特征修正方法能够提高语义分割精度，该方法在Semantic3D数据集上的语义分割精度达到78.7%，较RandLA-Net方法提升1.3%，且在保证分割精度的前提下其点云预处理计算复杂度和网络计算时间明显降低，在处理点数为10万~100万的大规模点云时，点云语义分割速度较SPG、KPConv等方法提升2~4倍。

Point Cloud Semantic Segmentation Method Based on Block Feature Fusion

A semantic segmentation method integrating block features is proposed to reduce the computational complexity of semantic segmentation of multi-class 3D objects in outdoor large-scale point cloud scenes.The square grid segmentation method is used to divide, sample, and combine the 3D point cloud, obtain the simplified point cloud combination block set, and input it into the block feature extraction and fusion network, to obtain the feature correction vector of each block.In order to correct the wrong feature caused by segmentation, the global feature correction network of the point cloud block is designed, and the feature correction vector is fused with the global feature of the original point cloud in the form of a residual.Based on this, the square mesh segmentation size is introduced into the back-propagation process as a parameter of the neural network for optimization, to establish an efficient point cloud semantic segmentation network.The experimental results show that the back propagation algorithm can optimize the segmentation size to near the optimal value, and the global feature correction method in the proposed network can improve the semantic segmentation accuracy.The semantic segmentation accuracy of this method on the Semantic3D dataset is 78.7%, which is 1.3% higher than the RandLA-Net method, and its point cloud preprocessing computational complexity and network computing time are significantly reduced on the premise of ensuring segmentation accuracy.When processing large-scale point clouds with 100 000‒1 million points, the semantic segmentation speed of point clouds is 2‒4 times higher than that of SPG, KPConv, and other methods.