Deep-Broad Learning Network for Semantic Segmentation of 3D Point Clouds

Wei Song, Tiancong Yan, Weilin Li, Kyungeun Cho, Ying Li

Abstract: 3D point cloud semantic segmentation provides effective technical support for environmental understanding and analysis in systems such as autonomous vehicles. To achieve high-precision 3D point cloud semantic segmentation, this paper proposes a 3D point cloud semantic segmentation algorithm based on the Deep-Broad Learning Network (DBLNet). First, the input to DBLNet is raw point cloud data. After passing through a dynamic graph convolution feature extraction module optimized by the Graph Sample and Aggregation (GraphSAGE) algorithm, the point cloud is transformed into high-dimensional features. Next, the semantic features of the point cloud are processed by a shared broad learning system to obtain the final semantic segmentation results. In DBLNet, the front multiple layers of the feature extraction module remain unchanged, while the deep edge convolutional network in the feature extraction module is replaced by a graph learning network using the GraphSAGE method. The GraphSAGE layer updates point features, reducing edge feature computations and enhancing feature extraction for complex point clouds. DBLNet replaces the fully connected network in traditional BLS with a shared perceptron, reducing layers and improving segmentation accuracy. The segmentation Mean Intersection over Union (mIoU) of DBLNet on the S3DIS dataset reaches 60.4%.