Adaptive Geometric Feature Learning Network-Based Point Cloud Segmentation for Radius Measurement of Grid-Stiffened Cylindrical Shell

Radius measurement of grid-stiffened cylindrical shell is a significant task for ensuring the welding quality of launch propellant tank. With the widespread adoption of 3-D laser scanning technology, cylinder fitting based on scanned point cloud of shell shapes has been providing an effective means for accurate radius measurement. To overcome the decrease in measurement accuracy caused by stiffener parts on the cylindrical shell, an adaptive geometric feature learning network (AdaGFLNet) featured with the flattening projection (FP) block and the neighbor attention (NA) block is designed to segment the skin points and stiffener points from the point cloud. The FP block captures the discriminative features for exploring global geometric differences via the projection representation. The NA block achieves local feature correlation learning from similar geometric structures with the squeeze-and-excitation attention mechanism. With the extracted skin points, the cylinder fitting algorithm is further employed to realize radius measurement. In comparative experiments, the results demonstrate the effectiveness of the proposed method for point cloud segmentation of skin points and stiffener points, and further for radius measurement of grid-stiffened cylindrical shell over other ones.