自动驾驶 3D 目标检测研究综述
作者:
作者单位:

北京工业大学

作者简介:

通讯作者:

中图分类号:

U463.6;TP391.41

基金项目:

国家重点基础研究发展计划(973计划),国家自然科学基金项目(面上项目,重点项目,重大项目)


3D Object Detection Algorithms in Autonomous Driving: A Review
Author:
Affiliation:

Beijing University of Technology

Fund Project:

The National Key Technologies R&D Program of China,The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    精确实时地进行目标检测是自动驾驶车辆能够准确感知周围复杂环境的重要功能之一,如何对周围物体的尺寸、距离、位置、姿态等3D信息进行精准判断是自动驾驶3D目标检测的经典难题.服务于自动驾驶的3D目标检测成为近年来炙手可热的研究领域,本文对该领域主要研究进展进行了综述.首先,介绍自动驾驶感知周围环境各相关传感器的特点.其次,介绍了3D目标检测算法并按照传感器获取数据类型将其分为:基于单目/立体图像的算法、基于点云的算法以及图像和点云融合的算法.接下来,对每类3D目标检测的经典算法以及改进算法进行了详细综述、分析、比较,梳理了当前主流自动驾驶数据集及其3D目标检测算法的评估标准,并对可见文献广泛采用的KITTI和NuScenes数据集实验结果进行对比及分析,归纳了现有算法存在的难点和问题.最后,提出自动驾驶3D目标检测在数据处理、特征提取策略、多传感器融合和数据集分布问题方面可能遇到的机遇及挑战,并对全文进行总结及展望.

    Abstract:

    Accurate and real-time object detection is one of the important functions for autonomous vehicles to accurately perceive the surrounding complex environment. Nevertheless, how to get the accurate size, distance, position, posture and other 3D information of surrounding objects is a classic problem. 3D object detection for autonomous driving has become a popular research field in recent years. Main research progress in this field is reviewed. Firstly, the characteristics of relevant sensors in the surrounding environment of autonomous driving is introduced. Secondly, the development of object detection from 2D to 3D is introduced and the loss functions is applied for optimization. According to the type of data acquired by the sensor, 3D object detection algorithms is categorized into three types, there are algorithms based on monocular/stereo images, point clouds, image and point cloud fusion. Futher more, the classic and improved algorithms for each type of 3D object detection are reviewed, analyzed, and compared in detail. Simultaneously,the mainstream autonomous driving datasets and the evaluation criteria of their 3D object detection algorithms are summarized. Extensive experiment results of KITTI and NuScenes datasets are also compared and analyzed, which is widely used inpresent literature, summarizing the difficulties and problems of the existing algorithms. Besides, data processing, feature extraction strategy, multi-sensor fusion and data distribution problems are proposed, which could be able to meet the opportunities and challenges of 3D object detection, in hope of inspiring more future work.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2022-04-14
  • 最后修改日期:2023-01-16
  • 录用日期:2022-09-03
  • 在线发布日期: 2022-09-17
  • 出版日期: