智能科学与技术学报 ›› 2023, Vol. 5 ›› Issue (1): 7-31.doi: 10.11959/j.issn.2096-6652.202312

• 综述与展望 • 上一篇    下一篇

3D目标检测方法研究综述

黄哲1, 王永才1,2, 李德英1   

  1. 1 中国人民大学信息学院,北京 100872
    2 警务物联网应用技术公安部重点实验室,北京 100048
  • 修回日期:2023-03-01 出版日期:2023-03-15 发布日期:2023-03-01
  • 作者简介:黄哲(1996- ),女,中国人民大学信息学院博士生,主要研究领域为计算机视觉、3D 目标检测
    王永才(1978- ),男,博士,中国人民大学信息学院副教授、博士生导师,主要研究领域为物联网、智能感知、网络定位、视觉感知、惯导融合定位
    李德英(1965- ),女,中国人民大学信息学院教授、博士生导师,主要研究领域为物联网、智能网络算法与分析
  • 基金资助:
    国家自然科学基金资助项目(61972404);国家自然科学基金资助项目(12071478);国家自然科学基金资助项目(61732006)

A survey of 3D object detection algorithms

Zhe HUANG1, Yongcai WANG1,2, Deying LI1   

  1. 1 School of Information, Renmin University of China,Beijing 100872, China
    2 Key Laboratory of Police Internet of Things Application Ministry of Public Security, Beijing 100048, China
  • Revised:2023-03-01 Online:2023-03-15 Published:2023-03-01
  • Supported by:
    The National Natural Science Foundation of China(61972404);The National Natural Science Foundation of China(12071478);The National Natural Science Foundation of China(61732006)

摘要:

3D 目标检测是自动驾驶、虚拟现实、机器人等应用领域的重要基础问题,其目的是从无序点云中框取出描述目标最准确的3D框,例如紧密包围行人或车辆点云的3D框,并给出目标3D框的位置、尺寸和朝向。如今,基于双目视觉、RGB-D相机、激光雷达构建的纯点云的3D目标检测,融合图像和点云多模态信息的3D目标检测,是两类主要的方法。首先介绍了3D点云的不同表示形式和特征提取方法,然后从传统机器学习类算法、非融合深度学习类算法、基于多模态融合的深度学习类算法3个层面,逐层递进地介绍各类3D目标检测方法,对类别内部和各类之间的方法进行分析和对比,深入分析了各类方法之间的区别和联系,最后论述了3D目标检测仍存在的问题和可能的研究方向,并对3D目标检测研究的主流数据集和主要评价指标进行了总结。

关键词: 深度学习, 3D目标检测, 多模态融合, 点云, 自动驾驶

Abstract:

3D object detection is a fundamental problem in autonomous driving,virtual reality,robotics,and other applications.Its goal is to extract the most accurate 3D box characterizing interested targets from the disordered point clouds,such as the closest 3D box surrounding the pedestrians or vehicles.The target 3D box's location,size,and orientation are also output.Currently,there are two primary approaches for 3D object detection: (1) pure point cloud based 3D object detection,in which the point clouds are created by binocular vision,RGB-D camera,and lidar; (2) fusion-based 3D object detection based on the fusion of image and point cloud.The various representations of 3D point clouds were introduced.Then representative methods were introduced from three aspects: traditional machine learning techniques; non-fusion deep learning based algorithms; and multimodal fusion-based deep learning algorithms in progressive relation.The algorithms within and across each category were examined and compared,and the differences and connections between the various methods were analyzed thoroughly.Finally,remaining challenges of 3D object detection were discussed and explored.And the primary datasets and metrics used in 3D object detection studies were summarized.

Key words: deep learning, 3D object detection, multimodal fusion, point cloud, autonomous driving

中图分类号: 

No Suggested Reading articles found!