2023 年 89 巻 1 号 p. 70-76
Nowadays, it is possible to capture the whole surface areas of buildings in three-dimensional point clouds using laser scanners. However, it is still challenging to find the damaged areas of the building data automatically. The objective of the study is to investigate the use of deep neural networks in combination with the building data captured by the terrestrial laser scanner for damage detection. The post-seismic building data for our experiments is from a structural lab. To find out what kinds of DNNs are useful for damage detection, the deep neural networks we choose are PointNet and PointNet++ (MSG, MRG, and SSG). Since innovative DNNs can only directly operate on small 3D rigid models, we transform the detection problem into a classification problem by voxelizing the whole façade before feeding into the 3D DNNs. Meanwhile, we find a pair of optimal parameters after investigating the impact of different point sizes and voxel sizes on detection results. By comparison, we perform a point-based classification by Random Forest to highlight the effectiveness of our approach.