[Paper Review] A comprehensive survey on point cloud registration

Xiaoshui Huang, Guofeng Mei+2 more|arXiv (Cornell University)|Mar 3, 2021

Robotics and Sensor-Based Localization116 references184 citations

TL;DR

This survey reviews optimization-based and deep learning methods for same-source and cross-source point cloud registration, introduces a cross-source benchmark, and discusses applications and future directions.

ABSTRACT

Registration is a transformation estimation problem between two point clouds, which has a unique and critical role in numerous computer vision applications. The developments of optimization-based methods and deep learning methods have improved registration robustness and efficiency. Recently, the combinations of optimization-based and deep learning methods have further improved performance. However, the connections between optimization-based and deep learning methods are still unclear. Moreover, with the recent development of 3D sensors and 3D reconstruction techniques, a new research direction emerges to align cross-source point clouds. This survey conducts a comprehensive survey, including both same-source and cross-source registration methods, and summarize the connections between optimization-based and deep learning methods, to provide further research insight. This survey also builds a new benchmark to evaluate the state-of-the-art registration algorithms in solving cross-source challenges. Besides, this survey summarizes the benchmark data sets and discusses point cloud registration applications across various domains. Finally, this survey proposes potential research directions in this rapidly growing field.

Motivation & Objective

• Provide a thorough overview of same-source point cloud registration methods (optimization-based, feature-learning, and end-to-end learning).
• Summarize cross-source registration challenges and existing solutions.
• Explain connections between optimization-based and deep learning approaches.
• Introduce a cross-source benchmark to evaluate state-of-the-art registration algorithms.
• Discuss applications and propose future directions for point cloud registration.

Proposed method

• Survey literature from 1992–2021 across optimization-based, feature-learning, and end-to-end registration paradigms.
• Classify methods into same-source and cross-source categories and into ICP-based, graph-based, GMM-based, and SDP-based approaches.
• Analyze advantages, limitations, and convergence properties of each category.
• Present a new cross-source benchmark to evaluate registration methods under cross-sensor challenges.
• Synthesize connections between traditional optimization and modern deep learning techniques.

Experimental results

Research questions

• RQ1 What are the key challenges and techniques in same-source versus cross-source point cloud registration?
• RQ2 How do optimization-based methods compare with deep learning approaches in terms of robustness and efficiency?
• RQ3 What is the nature of the connections between optimization strategies and deep learning methods for registration?
• RQ4 How does cross-source registration differ from same-source registration, and how can benchmarks capture this gap?
• RQ5 What are the open questions and future directions for point cloud registration research?

Key findings

• Comprehensive coverage of both traditional and deep learning registration methods for same-source data.
• Summary of cross-source registration challenges and the proposed benchmark to evaluate state-of-the-art methods.
• Identification of connections and complementarities between optimization-based and deep learning approaches.
• Discussion of applications across domains and the need for future research directions in cross-source scenarios.
• Highlighting of scalability and efficiency issues in semi-definite relaxation and graph-based methods for larger point sets.