Semantic and volumetric 3D plant structures modeling using projected image of 3D point cloud

今渕 貴志; 川端 邦明

Proceedings of 2024 IEEE/SICE International Symposium on System Integration (SII2024) (Internet), p.141 - 146, 2024/01

http://dx.doi.org/10.1109/SII58957.2024.10417376

This paper describes a method for volumetric-based semantic 3D modeling from 3D point cloud obtained in a plant environment. In order to calculate a radiation dose distribution of the workspace in decommissioning, the shape, arrangement, materials, and thicknesses of structures are essentially required in addition to dose values. However, it is costly to create such enriched 3D models from 3D point cloud. In this study, we propose a method to create 3D models with structural category and material thickness by combining 2D image-based deep learning and volumetric reconstruction method. To discriminate structures, structural category labels are predicted by a pre-trained 2D semantic segmentation network on projected image created from 3D point cloud. Then, a triangular mesh is generated from the integrated Truncated Signed Distance Function (TSDF) according to prediction labels. In addition, we optimize the TSDF thickness assignment function to reduce surface distance error. Our evaluation reports thickness and surface distance errors when generating meshes with three different structural categories in a mock-up plant environment.