Weld bead segmentation using RealSense depth camera based on 3D global features and texture features of subregions

Weld bead segmentation is important for the automation of quality control and inspection of welded joints. Due to the irregularity and the lack of topological information of the point cloud, it is hard to effectively segment the bead from the weld in 3D point cloud. This research investigates a novel method by using the 3D texture features and global features of point cloud subregions to achieve submillimeter precision segmentation of 3D weld point clouds captured by a RealSense depth camera. First, a preprocessing algorithm that combines a 2D image processing method with a segmentation method based on the global bead profile is proposed to improve the segmentation performance of the algorithm. Then, based on different feature selection algorithms and machine learning (ML) classifiers, the classification performance of the algorithm is effectively optimized. The experimental results show that choosing appropriate subregion sizes and features can significantly improve the segmentation performance of the algorithm. The average error in the bead width is less than 1 mm. Compared to other methods based on laser profile and ML, it has much better accuracy.