Cross-source point cloud registration network with low overlap based on multi-scale features and attention mechanisms

Machine vision-based detection methods have been widely applied in the detection of aircraft skin damage. During drone inspection processes, a key step is to spatially locate high-resolution detailed images of aircraft skin from multiple angles onto a three-dimensional point cloud model of the aircraft. This relies on the rigid registration of image center position coordinate point cloud with the aircraft 3D point cloud. To address the issues of low accuracy and poor robustness encountered by existing registration algorithms when dealing with heterogeneous point clouds with significant differences in density and low overlap, this paper presents a novel cross-source point cloud registration network. The network integrates multi-scale information from the point cloud and employs an attention mechanism to identify representative overlapping points. First, the network achieves initial correspondences using the multi-scale geometric features and positional information of the point cloud. Then, an overlapping feature guidance module predicts the overlapping score of the point cloud. By utilizing information interaction through the attention mechanism, the network combines point overlapping scores with fused features to filter out representative overlapping points, achieving precise correspondences in the point cloud. The network employs weighted singular value decomposition (SVD) to estimate two sets of transformation matrices, yielding the relative pose parameters of the point cloud. Experiments were conducted in an unsupervised manner. The experimental results on the ModelNet40 dataset and the aero object dataset aircraft measurement data showed that, compared to other existing traditional and learning-based methods, this approach demonstrated excellent performance in terms of registration accuracy and robustness.