基于多帧信息的多传感器融合三维目标检测  

Multi-Sensor Fusion 3D Object Detection Based on Multi-Frame Information

作　　者：吴绍斌[1] 耿家琳 吴超[1] 闫泽新 陈恺宇 WU Shaobin;GENG Jialin;WU Chao;YAN Zexin;CHEN Kaiyu(School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京理工大学机械与车辆学院,北京100081

出　　处：《北京理工大学学报》2023年第12期1282-1289,共8页Transactions of Beijing Institute of Technology

基　　金：区域创新发展联合基金(U19A2083)。

摘　　要：为提高三维目标检测中多传感器融合的效果,并利用前后帧之间的特征关联,提高目标检测的准确率,提出了一种基于多帧信息的多传感器特征融合三维目标检测网络.首先通过基于指导点的特征映射模块,将图像相机视角特征转换为鸟瞰图特征,并通过自适应融合模块对点云特征和图像特征进行融合;之后利用历史帧跟踪信息,融合多帧特征;最后采用基于CenterPoint检测头进行目标检测.在nuScenes数据集和实车上对三维目标检测网络进行了测试,试验结果表明该网络具有更高的精度和实时性.In order to improve the effectiveness of multi-sensor fusion in 3D object detection and improve the accuracy of object detection in the utilization of the feature correlation between the front and back frames,a multisensor feature fusion 3D object detection network based on multi frame information was proposed.Firstly,using a feature mapping module based on guidance points to convert the camera perspective features of the image into aerial features,the point cloud features and image features were fused with an adaptive fusion module.Afterwards,utilizing historical frame tracking information,multiple frame features were fused.Finally,a detection head CenterPoint was used to detect the objects and to test the 3D object detection network with a dataset nuScenes and real vehicles.The experimental results show that the network can provide higher accuracy and realtime performance.

关 键 词：目标检测 多传感器融合 多帧特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]