Point Cloud Registration in 3D Reconstruction: A Survey of Traditional Geometric Methods, Deep Learning, and System-Level Integration

Abstract

In scenarios involving multi-source asynchronous data acquisition, point cloud registration is often constrained by limited overlap, geometric information loss due to occlusion, noise, non-uniform sampling, repetitive structures, and discrepancies between sensing sources. This survey focuses on rigid point cloud registration, first clarifying the key elements defining the problem, including data conditions, task formulations, and evaluation criteria. It then reviews traditional geometric methods, analyzing how they construct local objectives, establish correspondence criteria, and ensure global consistency. The article also explores learning-based methods from various perspectives, highlighting representation learning, overlap modeling, relational reasoning, knowledge transfer, and integration into complete reconstruction systems. Furthermore, it outlines the applications of pairwise registration outputs in techniques such as pose-graph optimization, surface fusion, NeRF, and 3D Gaussian Splatting. Finally, the survey discusses unresolved challenges regarding reliability, generalization, efficiency, and system integration..