A Graph-Voxel Joint Convolution Neural Network for ALS Point Cloud Segmentation

Abstract

A deep convolution neural network is frequently used in airborne laser scanning (ALS) point cloud segmentation. In this study, we propose a joint graph-voxel convolution network to recognize on-ground objects accurately. In our network, the spatial context features of an input point cloud are first extracted by a designed U-Net via sparse convolution neural networks (SU-Net). Next, the extracted features are used as input features of a designed graph convolution network. We design a graph convolution module called the G-Net to extract the local spatial structure of each point. To enhance the representation of spatial context information, we initialize a graph based on a horizontal direction to enhance the difference between objects and the ground. The output probabilities of SU-Net and the graph convolution model are weighted as the input of the conditional random field optimizing model. The framework proposed in this study exhibits high processing efficiency. Experiments on the semantic 3D labeling dataset of the International Society for Photogrammetry and Remote Sensing (ISPRS) and 2019 IEEE Data Fusion Contest Dataset demonstrate that the proposed model significantly improve the highest F1 score and outperforms various previous networks.

Keywords

• ALS
• point cloud
• deep learning
• graph
• sparse convolution
• conditional random field