This study proposes a method for automatically recognizing structural members in the sub-assembly process of shipyards using 3D point cloud data acquired by a scanner, as part of welding automation efforts. In the sub-assembly process, many members ha...
This study proposes a method for automatically recognizing structural members in the sub-assembly process of shipyards using 3D point cloud data acquired by a scanner, as part of welding automation efforts. In the sub-assembly process, many members have similar or symmetrical shapes and sizes, making it difficult for conventional AI models or vision-based methods to accurately distinguish between them. To address this issue, an algorithm was developed to map the measured data to the corresponding CAD data.
The measured data often contains unnecessary elements such as the base plate and noise. Therefore, before placing the members, the base surface is scanned separately to establish a reference plane, which is then used to remove the floor data. Next, a grid-based filtering process removes sparse or discontinuous points and segments the data so that closely positioned members are not connected as one. From each segmented member, boundary and stiffener points are extracted and used for subsequent registration. Since mapping all CAD data to each scanned member is time-consuming, candidate CAD models are first selected based on several criteria, including area, bounding box, histogram, and hole grid maps. To distinguish between members with similar areas but different shapes, additional features such as bounding box aspect ratio are used. For more precise comparison, histogram and hole grid maps are calculated to further differentiate CAD models from the scanned parts. This stepwise comparison significantly reduces the number of CAD candidates, allowing detailed registration only for the remaining few and thereby reducing computational costs. Final registration is performed using the Iterative Closest Point (ICP) algorithm for boundary and stiffener points. Considering that CAD data are represented by lines and scanned data by points, ICP is applied iteratively with multiple initial positions to avoid local minima, and the result with the smallest alignment error is selected. For cases where a sufficient mapping ratio is achieved, the error and mapping ratio are recorded, and the CAD model with the smallest final error is determined as the corresponding member. The proposed method was validated using various real-world datasets, demonstrating its effectiveness in practical shipyard environments.