An improved deep learning based pointnet++ architecturefor semantic segmentation of lidar point clouds

PointNet++, a 3D Deep Neural Network (DNN) architecture that can directly process point clouds and is designed for indoor point clouds, is used for the semantic segmentation of Airborne Laser Scanning (ALS) point clouds. Initially, the PointNet++ architecture is adapted to large-scale airborne LiDAR point clouds for multi-class land cover and binary ground/non-ground semantic segmentation tasks. Two new approaches, called Adaptive 1 and Adaptive 2, are proposed to make the architecture more robust against varying point densities by adapting to point density instead of using the original bacthing strategy. Furthermore, a new hybrid model that combines different grouping layers of the architecture is developed and evaluated in terms of processing time and semantic segmentation performance. In addition, an approach is proposed to automate the selection of the initial value for the neighborhood search radius. Also, the effects of additional raw attributes such as intensity, return number and number of returns on the performance of the architecture are investigated. The semantic segmentation performances of the proposed approaches are evaluated on ISPRS Vaihingen, DALES, irregular DALES, and OpenGF point cloud datasets. Specifically, on the OpenGF test datasets with high point density variation, the architecture's weighted average Intersection over Union (mIoU) accuracy has improved by up to 17% with the Adaptive 1 approach and 11% with the Adaptive 2 approach. Adaptive 1 and Adaptive 2 approaches have achieved performance increases of 1% and 3% in mIoU accuracy on the irregular DALES dataset, respectively. The results show that the Adaptive 1 and Adaptive 2 approaches improve the generalization ability of the PointNet++ architecture for mIoU by approximately 3% and 5%, respectively. These proposed approaches significantly reduce the need for trial-and-error parameter selection while making the architecture more robust against both point sampling and parameter selection.