A point clouds registration model with detail features capture and geometric self-attention for complex workpieces with low coincidence point clouds

Abstract Point clouds registration is one of the keys to measuring workpiece size, because the three-dimensional contours of workpieces are constructed from point clouds captured at different angles by using one or several laser cameras. A point clouds registration model with detail features capture and geometric self-attention for complex workpieces with low coincidence point clouds is presented in this paper. Under the ROPNet model framework, the DGCNN and PnP-3D network are combined to extract point cloud features. A geometric self-attention (GS) mechanism is incorporated into the transformer-based feature matching removal (TFMR) model to elevate registration accuracy of complex workpieces with low overlapping point clouds. The simulation verification results based on workpieces point cloud datasets with noise-free and noisy conditions indicate that the proposed method surpasses ICP, Go-ICP, PCRNet, PRNet, and ROPNet. Especially, compared to ROPNet, the given model achieves a notable reduction in isotropic rotation error Err(R) by 3.1704 and translation error Err(t) by 0.04. Moreover, the model also can maintain high registration accuracy for workpieces with low-integrity point clouds.