Damage Detection of the RC Building in TLS Point Clouds Using 3D Deep Neural Network PointNet plus

We are working on a research project to evaluate the safety and structure of reinforced concrete buildings damaged by earthquakes using point cloud data acquired by a terrestrial laser scanner. We propose a framework for damage analysis as a classification problem that divides a building in point clouds into small 3D voxel grids and determines which voxels are damaged, instead of detecting the damaged parts from the point cloud of the whole building. Our framework is divided into three steps: First, the damaged building in point clouds is divided into small 3d voxel grids. Second, every voxel is fed into the deep neural network for damage classification. As a deep neural network to classify the voxel grids, we used PointNet++. Finally, the original damage map is refined by a simple cluster analysis. After post-processing, the recall reaches 0.929. That is, 92.9% of damage portions are correctly detected although the damage map still contains some FPs.