PointNet++ Network Architecture with Individual Point Level and Global Features on Centroid for ALS Point Cloud Classification

Airborne laser scanning (ALS) point cloud has been widely used in the fields of ground powerline surveying, forest monitoring, urban modeling, and so on because great convenience it brings to people’s daily life. However, sparsity uneven distribution clouds increases difficulty setting uniform parameters for semantic classification. The PointNet++ network is an end-to-end learning irregular data highly robust small perturbations input points along with corruption. It eliminates need calculate costly handcrafted features provides a new paradigm 3D understanding. each local region output abstracted by its centroid feature that encodes centroid’s neighborhood. learned may not contain relevant information itself random sampling, especially large-scale neighborhood balls. Moreover, point’s global-level sample layer also marked. Therefore, this study proposed modified architecture which concentrates point-level global towards facilitate approach utilizes Focal Loss function solve extremely category ALS clouds. An elevation- distance-based interpolation method objects exhibit discrepancies elevation distributions. experiments Vaihingen dataset International Society Photogrammetry Remote Sensing GML(B) demonstrate additional contextual support classification achieves high accuracy simple discriminative models state-of-the-art performance power line categories.