面向3D目标检测任务的数据增强方法研究进展

Research Progress of Data Augmentation Methods for 3D Object Detection

摘要: 基于深度学习的3D点云目标检测技术在自动驾驶、智慧工业等领域快速发展的过程中起到了关键性及支撑性作用。然而，由于3D点云覆盖空间广阔、数据稀疏的特点，为了实现更高精度的目标检测，需要对原始点云数据进行数据增强操作。目前，针对2D图像数据增强方法的研究较为广泛，但是面向3D点云数据的增强方法研究仍处于早期阶段。因此，本文旨在针对3D目标检测数据增强方法研究进展进行综述，首先介绍了3D目标检测的基本技术和流程，然后介绍并分析了面向3D目标检测任务的数据增强方法，具体分为三个类别，包括基于2D图像衍化而来的3D点云数据增强方法、针对3D点云设计的增强方法以及混合与创新型数据增强方法。最后讨论了该领域存在的挑战以及未来的发展方向，为未来该领域的研究人员提供参考。

Abstract: Deep learning-based 3D point cloud object detection technologies have played a crucial and supportive role in the rapid development of fields such as autonomous driving and smart industry. However, due to the vast coverage and sparse nature of 3D point clouds, data augmentation operations are necessary to achieve higher precision in object detection. Currently, there is extensive research on data augmentation methods for 2D images, but the study of augmentation methods for 3D point cloud data is still in its early stages. Therefore, this paper aims to provide a comprehensive review of the progress in data augmentation methods for 3D object detection. It first introduces the basic techniques and processes of 3D object detection, then presents and analyzes data augmentation methods for 3D object detection tasks, which are divided into three categories: methods derived from 2D image augmentation applied to 3D point clouds, methods designed specifically for 3D point clouds, and hybrid and innovative data augmentation methods. Finally, the paper discusses the challenges in this field and future directions for development, offering a reference for researchers in this area moving forward.