Abstract:
Point cloud semantic segmentation is a key technique for the digital twin construction of water conservancy projects, which can realize the identification and change detection of terrain features. However, constructing full-range point cloud data of the water conservancy environment remains one of the critical challenges in achieving comprehensive digital twin construction. Meanwhile, the openness of the water conservancy environment also makes its point cloud data structure highly complex, posing challenges to the accuracy and robustness of point cloud semantic segmentation algorithms. Therefore, we adopt unmanned aerial vehicle (UAV)-borne lidar to scan water conservancy scenes and construct a large-scale point cloud dataset, Water Conservancy Segment 3-D (WCS3D), with approximately 265 million points. On this basis, we propose a point cloud segmentation model named PointBiMssc based on a bidirectional multiscale attention mechanism for point cloud semantic segmentation in the water conservancy environment. A series of experiments conducted on the WCS3D dataset demonstrate that the PointBiMssc model can accurately complete point cloud semantic segmentation tasks and generate high-precision segmentation boundaries, outperforming the latest transformer-based models in mean intersection over union (mIoU) and overall pointwise accuracy (OA) evaluation metrics to achieve state-of-the-art performance. Code: https://github.com/JJBUP/PointBiMssc, dataset: https://github.com/XTU-SCS-HappyCV/WCS3D. Β© 2004-2012 IEEE.
