Fast Online Incremental Segmentation of 3D Point Clouds from Disaster Sites

In monitoring disaster sites, mobile robots represent a fast, reliable, and practical option to remotely inspect active areas in disaster management for applications such as risk management, search and rescue, and structural assessment purposes. Mobile robots can efficiently collect laser scan data and reconstruct the state of ongoing disaster relief in the form of 3D point clouds. Current point cloud processing methods are mostly designed to work as a post-processing step and are inefficient when applied in real-time. Additionally, object segmentation on point cloud data from the disaster sites is challenging due to data impurities and occlusions. To overcome these issues, this paper proposes an instance point cloud segmentation method that incrementally builds a 3D map for robotic scans of infrastructure. In the first step, the proposed neural network, named Dynamic Graph PointNet (DGPointNet), is trained to classify objects in the disaster environment while building up a semantic 3D map of the environment. Additionally, the proposed method predicts object instance labels by using a sequence of predicted semantic point cloud data. The proposed method shows strong performance over the state of art segmentation models in terms of semantic segmentation, instance segmentation, and processing time using point cloud data collected from a custom-built laser-scanning robot at an outdoor simulated disaster site.