The 3D point cloud is a crucial type of data collected by LiDAR sensors and widely used in transportation-related fields due to its concise description and accurate localization. To better capture the hierarchical representation of the point cloud, deep neural networks (DNNs) are employed to process considerable disorder 3D points and accomplish promising achievements in various computer vision tasks such as pedestrian detection and vehicle recognition. However, point cloud labeling is time-consuming and labor-intensive due to the disordered and sparse attribute of the point cloud. Therefore, self-supervised learning (SSL), an unsupervised training paradigm that mines effective information from the data itself without human annotations, is considered an essential solution for moving away from label dependency via its brilliant pre-training task design. This paper provides a comprehensive survey of recent advances in deep neural network-based point cloud SSL. We initially introduce the definition, motivation, general pipeline as well as background to describe the concept of SSL and its combination with point cloud in detail. In addition, we present an innovative taxonomy of point cloud SSL methods, separating current approaches into four broad categories based on the pretexts' characteristics. We then summarize and compare the performances of the proposed SSL methods in literature on multiple downstream tasks quantitatively and qualitatively on benchmark datasets. Finally, feasible future research directions are provided to alleviate current limitations of point cloud SSL and also enhance feature extraction capability and generalizability of pre-training models.

本文介绍了近期基于DNN的点云SSL的综述，包括定义、动机、背景、基于几种常见Task-agnostic预训练的创新分类方法，以及对于基准数据集的实验结果总结，最终给出了针对当前点云SSL限制的改进和未来研究方向。

点云数据的自监督学习：综述