Research on AI semantic segmentation model for large-scale point cloud data based on cloud-edge collaboration

Airborne LiDAR can rapidly acquire large-scale point cloud data, laying the foundation for AI semantic segmentation models, and showing increasingly important application value in areas such as topographic mapping, urban planning, disaster monitoring, and power patrol. For the point cloud classification task, existing studies have proposed deep learning models based on the PointNet architecture that can directly process raw point cloud data, while recent studies have explored graph convolutional neural networks based on point cloud topology. However, current models mostly focus on local geometric features and fail to fully exploit the global contextual associations of point clouds. In this paper, we propose a deep map attentional convolutional neural network (DACNN) designed to directly classify 3D point clouds. DACNN takes into account the spatial distribution relationships and characteristics between all points in the global. The deep attention module focuses on the local neighborhood of each point, dynamically generates convolutional weights based on spatial location and associates global spatial information. A series of comparative experiments demonstrated the effectiveness of our approach. Our method can be utilized in the cloud devices, where the predictive efficiency of the model after training is higher and can be deployed on edge devices.