首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 74 全文下载次数: 89
引用本文:

DOI:

10.11834/jrs.20243396

收稿日期:

2023-09-13

修改日期:

2024-05-08

PDF Free   EndNote   BibTeX
基于深度学习的成对点云刚性配准现状与进展
摘要:

点云刚性配准作为三维点云数据处理的一项基础任务,在自动驾驶、机器人、测绘遥感、医疗、工业设计以及文物保护等方面具有广泛应用。基于深度学习的点云配准方法能够自动学习高判别力的点云特征,取得较高的配准精度。为了便于后续研究学者对点云配准进行更深入有效的探索,本文对基于深度学习的成对点云刚性配准相关技术研究进行了系统的综述与分析。首先,本文介绍了基于深度学习的局部特征提取方法,然后介绍对应关系解算、姿态回归和场景流估计等三类基于深度学习的配准方法相关研究进展。接着,本文对现有的可用于点云刚性配准的公开数据集进行了系统的总结与归纳。最后,本文对点云配准研究现状进行了总结,并对未来相关研究方向进行展望。

Status and progress of deep learning-based pairwise point cloud rigid registration
Abstract:

Point cloud registration is the process of spatial alignment of two or more point clouds through geometric transformations. As a fundamental task in 3D point cloud data processing, it is an important preprocessing operation for tasks such as 3D modeling, object recognition and scene understanding. Due to the non-structural, sparse, and uneven characteristics of point cloud data, point cloud registration remains one of the hotspots and challenges in computer vision, mapping and remote sensing, although there have been a lot of studies. With the emergence and rapid development of neural networks, deep learning has shown great potential in applications such as point cloud classification, recognition, detection, and reconstruction. In recent years, many researchers have also attempted to apply deep learning techniques to point cloud registration. Deep learning-based point cloud registration methods can automatically learn highly discriminative and robust point cloud features that contain geometric structural information and semantic information, which are crucial for achieving high registration accuracy. In this paper, the research on pairwise point cloud rigid registration technology based on deep learning is systematically reviewed and analyzed. Firstly, the feature extraction network based on deep learning is introduced. Then, the research progress of the registration methods of correspondence estimation, pose regression and scene flow estimation is reviewed, and the characteristics, advantages and disadvantages of these three methods are summarized. Then, this paper systematically summarizes and categorizes existing publicly available datasets that can be used for rigid point cloud registration. Finally, this paper summarizes the current research status of point cloud registration, explains the advantages and limitations of existing methods in feature learning, registration accuracy, registration efficiency and other aspects, and prospects for future research directions, proposing three exploration directions: (1) Existing deep learning-based point cloud registration methods require a large amount of annotated data, while data annotation time and manpower cost are high. Methods such as few-shot learning, self-supervised learning, weakly supervised learning and unsupervised learning can alleviate the need in deep learning technology for large amounts of labeled data to some extent. (2) The matching primitives in existing deep learning-based point cloud registration methods still primarily focus on 3D points. Compared to geometric elements such as 3D line segments and planes, the ambiguity of 3D points is large and the mismatching rate is high. In the future, it is worth considering using deep learning techniques to automatically extract the geometric structural elements of scenes from 3D point cloud data, and then further solve the registration problem of 3D point clouds in large scenes based on these geometric structural elements. (3) Existing deep learning-based registration methods mainly utilize spatial geometric features of scenes, with less consideration of semantic information. In recent years, there have been rapid advances in deep learning-based point cloud semantic understanding techniques. Therefore, in the future, it is worth considering combining spatial geometric structural information with semantic information to solve the registration problem of complex scenes in 3D point clouds.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群