Point Cloud Registration

Point cloud registration estimates how two or more 3D scans relate, enabling map building, robot localization, scene reconstruction, and alignment of medical or industrial data. Classical methods such as ICP and GICP refine geometric matches, while RANSAC, optimal transport, and global optimization help handle outliers, partial overlap, and poor initial poses; learned descriptors and transformer-based matchers provide another route to correspondences. Current work combines these approaches with image, radar, inertial, semantic, or temporal information to support cross-sensor and deformable registration, including settings with sparse measurements. A central challenge is achieving reliable accuracy across unseen scenes and sensor setups while remaining efficient and robust to low overlap, ambiguity, and geometrically degenerate environments.