自动驾驶中基于深度学习的3D目标检测方法综述
Survey on deep learning-based 3D object detection methods in autonomous driving
梁振明 1黄影平 1宋卓恒 1丁建华1
作者信息
- 1. 上海理工大学,光电信息与计算机工程学院,上海 200093
- 折叠
摘要
随着激光雷达传感器和深度学习技术的快速发展,针对自动驾驶 3D目标检测算法的研究呈现爆发式增长.为了探究 3D目标检测技术的发展和演变,对该领域中基于深度学习的 3D检测算法进行了综述.根据车载传感器的不同,将当前基于深度学习的自动驾驶 3D目标检测算法分为基于相机RGB图像、基于激光雷达点云、基于RGB图像-激光雷达点云融合的 3D目标检测3 种类型.在此基础上,分析了各类算法的技术原理及其发展历程,并根据平均检测精度(mAP)指标,对比了它们的性能差异与模型优缺点.最后,总结和展望了当前自动驾驶 3D目标检测中仍然面临的技术挑战及未来发展趋势.
Abstract
With the rapid developments of LiDAR sensors and deep learning technology,researches on 3D object detection for autonomous driving had witnessed remarkable growth.In order to explore the evolution of 3D object detection technology in autonomous driving,the existing deep learning-based 3D object detection methods were summarized.Based on the rely-on sensors,these methods could be divided into three classes:Camera RGB image-based,LiDAR point cloud-based and RGB image-LiDAR point cloud fusion-based 3D object detection.On this basis,the methodology and chronological overview of different kinds of methods were analyzed,and their 3D detection performance and characteristics were compared according to the mean average precision(mAP)metrics.Finally,the principal technical challenges and potential development trends in future autonomous driving 3D object detection researches were summarized and explored.
关键词
3D目标检测/深度学习/自动驾驶/RGB图像/激光雷达点云/多传感器融合Key words
3D object detection/deep learning/autonomous driving/RGB image/LiDAR point cloud/multi-sensors fusion引用本文复制引用
基金项目
国家自然科学基金(62276167)
上海市自然科学基金(20ZR1437900)
出版年
2024
NETL
NSTL
维普
万方数据