*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: 3D Segmentation and Damage Analysis from Robotics Scans of Disaster Sites
Committee:
Dr. Yong K. Cho (Advisor), School of Civil and Environmental Engineering
Dr. Zsolt Kira, School of Interactive Computing
Dr. Frank Dellaert, School of Interactive Computing
Dr. Jun Ueda, School of Mechanical Engineering
Dr. Eric Marks, School of Civil and Environmental Engineering
Abstract: Mobile robots can be used to assist in post-disaster reconnaissance by traversing a disaster site and performing remote surveillance without incurring danger to human lives. Relief operations such as damage assessment, risk management, and search and rescue require an accurate 3D semantic map of the disaster site to be carried out safely and efficiently. Thus, there exists a research need to automatically identify building elements and detect structural damage from laser-scanned points clouds acquired by mobile robots. Current methods for point cloud semantic segmentation mostly perform direct class prediction at the point level without considering object-level semantics and generalizability across datasets. Moreover, current segmentation methods are lacking when applied to robotic real-time scanning because they are designed to operate on point clouds one at a time and do not incorporate information from new scans in an incremental fashion. This research proposes a learnable region growing method to perform class-agnostic point cloud segmentation in a data-driven and generalizable manner. In addition, an anomaly-based crack segmentation method is proposed where a deep feature embedding is used as a basis for separation between inlier and outlier points. Finally, an incremental segmentation scheme is used to process point cloud data in an online fashion and combine semantic information across multiple scans.
