ELMF-Net: Semantic Segmentation of Large-Scale Point Clouds via Efficient Local Feature Learner and Multiscale Fusion

The semantic segmentation of 3-D point clouds can precisely describe 3-D environmental information, serving as an important research direction for environmental perception in unmanned systems. However, existing methods face drawbacks owing to the limitations in local semantic feature representation and cross-scale information fusion capabilities. To address these issues, we propose ELMF-Net, an efficient and accurate semantic segmentation model for large-scale 3-D point clouds. First, we introduce a local feature learning method that does not rely on strict geometric relationships and establish a local feature learner (w-LFL) model to capture and aggregate locally semantic discriminative features from point clouds. Subsequently, a novel multiscale feature fusion (MSFF) module was designed to collaborate with the decoder to deeply integrate shallow encoding layer features at different resolutions and high-level semantic features from deep encoding layers, providing an efficient representation of objects with varying scales. Finally, we validate the performance of ELMF-Net on three large-scale datasets, Stanford large-scale 3D indoor spaces dataset (S3DIS), Toronto3D, and Semantic Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago (KITTI), demonstrating the excellent performance of the ELMF-Net network in large-scale, multitarget scene.