Cross-sensor adaptive semantic segmentation for mobile laser scanning point clouds based on continuous potential scene surface reconstruction

Semantic segmentation is a fundamental task for extracting road information from mobile laser scanning (MLS) point clouds. Recently, deep learning-based methods have shown superior performance in MLS point cloud semantic segmentation. However, MLS is usually equipped with different LiDAR sensors, which leads to point-level distribution differences in point clouds. Therefore, a deep network trained on the source domain point clouds often performs poorly on the target domain point clouds. In this paper, we propose a new cross-sensor adaptive semantic segmentation for MLS point clouds based on continuous potential scene surface reconstruction. Firstly, an implicit neural representation framework is introduced to reconstruct the continuous potential scene surface for MLS point clouds. Then, the source and target domain MLS point clouds are both transformed into a canonical domain based on the continuous potential scene surfaces to achieve point-level distribution alignment. Next, an adaptive neighbor vote strategy is designed to map the source domain training label to the canonical domain and map the canonical domain semantic segmentation results to the target domain. Three MLS point cloud datasets were used to evaluate the performance of the proposed method. The experimental results indicated that our approach can effectively achieve cross-sensor adaptive semantic segmentation for MLS point clouds. An implementation of the proposed method is available at: https://github.com/PCsFJNU/CrossSensorAdaptiveSemanticSeg.