基于点云的自动驾驶下三维目标检测
3D object detection in automatic driving based on point cloud
杨咏嘉 1钟良琪 1闫胜业1
作者信息
- 1. 南京信息工程大学 自动化学院,江苏南京 210044
- 折叠
摘要
针对当前三维目标检测算法对行人、骑行人等小目标检测效果不佳的缺点,提出一种改进PV-RCNN的三维目标检测算法.改进关键点下采样方式,通过滤除背景及离群点提高关键点在目标上的命中率;设计多尺度区域建议网络,尺度匹配的特征图提高边界框的生成质量;使用加入方向感知的DIoU损失函数优化边界框的回归.实验结果表明,与基准网络相比,算法在KITTI测试集的车辆、行人和骑行人的mAP分别提高了 0.77％、6.33％和2.05％,有效提高了网络性能.
Abstract
Aiming at the disadvantage of current 3D target detection algorithm for small targets such as pedestrian and cyclists,a 3D object detection algorithm based on improved PV-RCNN was proposed.The method of key points sampling was improved,and the hit rate of key points on the target was improved by filtering the background and outliers.A multi-scale region proposal network was designed.The feature map of scale matching improved the quality of proposal box generation.The DIoU loss func-tion with direction perception was used to optimize the regression of the bounding box.Experimental results show that compared with the benchmark network,the mAP of car,pedestrian and cyclist in the KITTI val set of the algorithm is improved by 0.77％,6.33％and 2.05％respectively,effectively improving the network performance.
关键词
深度学习/三维目标检测/特征金字塔/原始点云/交并比损失函数/特征融合/点云下采样Key words
deep learning/3D object detection/feature pyramid network/raw point cloud/IoU loss/feature fusion/point cloud downsampling引用本文复制引用
出版年
2024
NETL
NSTL
万方数据