CrackEmbed: Point feature embedding for crack segmentation from disaster site point clouds with anomaly detection

Laser-scanned point clouds can be used to represent the 3D as-damaged condition of building structures in a post-disaster scenario. Performing crack detection from the acquired point clouds is a critical component of disaster relief tasks such as structural damage assessment and risk assessment. Crack detection methods based on intensity or normals commonly result in noisy detections. On the other hand, deep learning methods can achieve higher accuracy but require a large dataset of annotated cracks. This research proposes an unsupervised learning framework based on anomaly detection to segment out cracked regions from disaster site point clouds. First, building components of interest are extracted from the point cloud scene using region growing segmentation. Next, a point-based deep neural network is used to extract discriminative point features using the geometry of the local point neighborhood. The neural network embedding, CrackEmbed, is trained using the triplet loss function on the S3DIS dataset. Then, an anomaly detection algorithm is used to separate out the points belonging to cracked regions based on the distribution of these point features. The proposed method was evaluated on laser-scanned point clouds from the 2015 Nepal earthquake as well as a disaster response training facility in the U.S. Evaluation results based on the point-level precision and recall metrics showed that CrackEmbed in conjunction with the isolation forest algorithm resulted in the best performance overall.