Spatial Feature Structure Generation Network for Airborne LiDAR Point Cloud Semantic Segmentation.

Area of Science:

• Computer Vision
• Geospatial Technology
• Artificial Intelligence

Background:

• Airborne Light Detection and Ranging (LiDAR) sensors generate point clouds widely used in surveying, intelligent monitoring, and autonomous driving.
• The irregular structure of point clouds presents significant challenges for segmentation, recognition, and understanding.

Purpose of the Study:

• To propose a novel network, SFSGNet, capable of effectively handling irregular point clouds.
• To improve the accuracy of point cloud semantic segmentation by generating expressive spatial structure features.

Main Methods:

• Developed a Spatial Feature Structure Generation (SFSG) module to fuse attribute and spatial coordinate features.
• Generated highly expressive spatial structure features for both semantic and spatial contexts.
• Evaluated SFSGNet on ISPRS 3D Labeling Benchmark, DALES, and OpenGF datasets.

Main Results:

• SFSGNet achieved an average F1 score of 74.4% and 85.4% overall accuracy on the ISPRS dataset.
• Achieved 89.3% average F1 score and 98.1% overall accuracy on the DALES dataset.
• Attained 98.2% average F1 score and 98.5% overall accuracy on the OpenGF dataset.

Conclusions:

• SFSGNet demonstrates excellent point cloud semantic segmentation performance.
• The proposed network shows strong generalization capabilities compared to other models.
• SFSGNet effectively addresses the challenges posed by irregular point cloud structures.