摘要
自动驾驶中传感器融合是感知系统的重要组成部分,雷达点云信息和视觉信息融合可以提高车辆的感知能力。然而现有的研究将雷达点投影到图像上时只是对雷达点简单的增加高度,无法提供更加准确的横向信息,缺乏空间信息。同时对两个模态只是进行简单的融合,虽然产生了一个联合表征,但不足以充分捕捉两种模态之间的复杂联系。文中同时增加了雷达点云的宽度来进行空间信息增强,另外设计了一种利用差异性特征注意力融合的方法,使两个模态进行跨模态交互融合。本文在具有挑战性的nuScenes数据集上对模型进行了评估,提出的模型的NDS评分和mAP分别达到了46.3%和33.9%,体现了优秀的性能。
Sensor fusion in autonomous driving is an important part of the perception system, and the fusion of radar point cloud information and visual information can improve vehicle perception.However, existing studies projecting radar points onto images simply add height to the radar points, which does not provide more accurate lateral information and lacks spatial information.Simultaneous fusion of the two modalities is only simple, which produces a joint representation but is not sufficient to fully capture the complex connection between the two modalities.In this paper, we simultaneously increase the width of the radar point cloud for spatial information enhancement, and additionally design a method for cross-modal interaction fusion of the two modalities using differential feature attention fusion.In this paper, the model is evaluated on the challenging nuScenes dataset, and the proposed model achieves 46.3% and 33.9% in NDS score and mAP,respectively, reflecting excellent performance.
作者
李艳
沈韬
曾凯
LI Yan;SHEN Tao;ZENG Kai(Yunnan Key Laboratory of Computer Technologies Application,Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming,Yunnan 650500,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2023年第1期26-33,共8页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61671225,61971208,61702128)
云南省应用基础研究计划项目重点项目(2018FA043)
云南省中青年学术技术带头人后备人才项目(Shen Tao,2018)
云南省万人计划青年拔尖人才项目(云南省人社厅(2018 73)资助项目。
关键词
差异性特征注意力
空间信息增强
跨模态融合
3D目标检测
differential feature attention
spatial information enhancement
cross-modal fusion
3D object detection
