Deep Learning for Point Cloud Registration: A Review of Core Architectures and Feature Extraction Strategies

Abstract

Point cloud registration aims to estimate the spatial transformation that aligns a source point cloud with a target point cloud, and constitutes a fundamental problem in three-dimensional reconstruction, robotic perception, autonomous driving, industrial metrology, and medical digitization. In recent years, deep learning has substantially reshaped the research paradigm of point cloud registration. The field has progressively shifted from traditional pipelines based on handcrafted geometric descriptors and local optimization toward data-driven frameworks that jointly learn feature representations, correspondence patterns, and rigid transformations. Focusing on the research question of which core architectures and feature extraction strategies are adopted by deep learning methods for point cloud registration, this review systematically analyzes the representative methodological routes developed in recent years. Specifically, we examine global feature regression architectures, correspondence learning architectures, coarse-to-fine hierarchical frameworks, and Transformer-based geometric attention models, and further discuss the evolution of feature learning from point-wise encoding to local geometric description, global context aggregation, multi-scale fusion, and cross-cloud interaction. On this basis, the review summarizes the intrinsic coupling among feature extraction, correspondence construction, and transformation estimation, compares the performance of representative models under a unified benchmark protocol, and outlines the main challenges and future directions in terms of generalization, efficiency, and geometric interpretability. This review is expected to provide a structured reference for subsequent studies on deep-learning-based point cloud registration.