Abstract: Semantic segmentation of three-dimensional point clouds is a key technology in the fields of autonomous driving and computer vision. Deep learning-based semantic segmentation methods suffer from inefficient operation due to their complex network structures. To address the inefficiency of deep neural networks in 3D point cloud semantic segmentation, Point-MLPBLS :Point cloud semantic segmentation network based on MLP cascaded broad learning system was proposed. First, The overall network was designed to adopt a dual network collaborative architecture: a feature extraction network and a point cloud segmentation network. Second, the feature extraction network was applied with sampling-grouping-aggregation operation to the input point cloud data and an inverted residual module was added to enhance local feature representation. Third, a feature propagation module was utilized to achieve upsampling, restores the original point cloud resolution, and performs multi-scale feature fusion to construct a refined point cloud representation. Finally, the point cloud segmentation network deep iterative training was replaced with a flattened multi-layer perceptron cascade structure. Spatially aware multi-layer perceptrons were embedded in the feature mapping layer to achieve high-precision inference. Experimental results show that Point-MLPBLS improves the mean Intersection-over-Union (mIoU) by 12.9 percentage points compared with PointNet++ on the S3DIS dataset, reduces the segmentation time by 41.1%, improves the efficiency of point cloud segmentation, and provides an efficient solution for 3D point cloud semantic segmentation.

Key words: 3D point cloud, semantic segmentation, broad learning system, feature fusion, attention mechanism

摘要： 三维点云语义分割是自动驾驶与计算机视觉领域的关键技术。基于深度学习的语义分割方法由于网络结构复杂，导致运行效率低下。针对深度神经网络在三维点云语义分割中效率低下的问题，提出一种基于多层感知机（MLP）级联宽度学习系统的点云语义分割网络 Point-MLPBLS。首先，整体网络采用双网络协同架构：特征提取网络和点云分割网络。其次，特征提取网络对输入的点云数据采用采样-分组-聚合操作，并追加倒置残差模块增强局部特征表达能力，再次，通过特征传播模块实现上采样，恢复原始点云分辨率并进行多尺度特征融合，构建精细化的点云表征。最后，点云分割网络采用扁平化多层感知机级联结构替代深层迭代训练，在特征映射层嵌入空间感知型多层感知机，实现高精度推理。实验结果表明，Point-MLPBLS在S3DIS 数据集上较 PointNet++的平均交并比(mIoU)提升了 12.9 个百分点，分割时间减少了 41.1%，提高了点云分割的效率，为三维点云语义分割提供了高效解决方案。

关键词: 三维点云, 语义分割, 宽度学习系统, 特征融合, 注意力机制