Registration of Nonuniform LiDAR Scans With Large Pose Variations for Aircraft Assembly

To ensure aircraft assembly quality, a reliable measurement system for aircraft pose estimation and adjustment is essential. Conventional laser tracker systems are limited to measuring only a few points and struggle to capture aircraft component deformation. Recent advancements in LiDAR systems offer rapid acquisition of millions of points, capturing complete surfaces and detecting deformation through precise registration. However, the non-uniform distribution, large pose, and non-distinctive features of aircraft scans pose challenges to LiDAR scan registration. To address these challenges, we propose a framework named Density and Pose Consistency Network (DPCNet), which extracts hierarchical features of aircraft scans. We design consistency losses based on LiDAR scan characteristics to improve registration accuracy. To evaluate registration methods on aircraft scans, we collect a LiDAR-scanned Aircraft Dataset (LAD). The LAD consists of multi-view LiDAR scans of aircraft parts, annotated with high-quality ground truth. We also perform data processing and augmentation to improve the quality and diversity of LAD. Experimental results show that DPCNet achieves satisfactory registration accuracy on the LAD, outperforming state-of-the-art conventional and learning-based methods. We further conduct experiments on ModelNet40 and KITTI datasets, demonstrating the effectiveness and scalability of our DPCNet in different scenarios.