Local and Global Structure for Urban ALS Point Cloud Semantic Segmentation With Ground-Aware Attention

Abstract

Interpretation of airborne laser scanning (ALS) point clouds plays a notable role in geoinformation production. As a critical step for interpretation, accurate semantic segmentation can considerably broaden various applications of ALS data. However, most existing methods cannot provide precise annotations and high robustness due to occlusions, varied point densities, and complex and incomplete object structures. Therefore, we developed a semantic segmentation framework focusing on ALS point clouds. The framework comprises contextual feature extraction from a local neighborhood, scene-aware global information representation, and a ground-aware attention module. To verify its effectiveness, comprehensive experiments were conducted on three airborne light detection and ranging (LiDAR) datasets: DublinCity, Dayton Annotated LiDAR Earth Scan (DALES), and DFC2019 datasets. The experimental results demonstrate that the proposed method achieves better segmentation performance than that some advanced methods. For the DublinCity dataset, our model&#x2019;s overall accuracy (OA) can be improved to 67.5&#x0025; with an average F₁ (Avg<inline-formula> <tex-math notation="LaTeX">$F_{1}$ </tex-math></inline-formula>) of 37.6&#x0025;. For the DALES dataset, our method achieved an OA of 96.5&#x0025; and a mean intersection over union (mIoU) of 77.6&#x0025;. Our method also achieves a more accurate result on the DFC2019 dataset than that obtained using other models with an OA of 94.8&#x0025; and an AvgF₁ of 81.4&#x0025;.