基于BEV多模态特征融合的三维目标检测算法

打开文本图片集

关键词：BEV特征；数据预处理；全局注意力融合；KITTI数据集；三维目标检测中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2025）12-038-3823-09doi：10.19734/j.issn.1001-3695.2025.03.0101

3D object detection algorithm based on BEV multimodal feature fusion

Li Xun1，Zhang Youbing1，2†，Zhou Kui1，Fu Yongzhi1，Song Jie1 （1.Sring-XJntbolfeltc，ubeo ShiyanHubei 4420oo，China；2.Jingchu Universityof Technology，Jingmen Hubei448000，China）

Abstract：Toaddress the problemof isuffcientdetectionaccuracyforsmalltargetobjectsinmultimodalfusionalgorithms， this paper proposeda3DobjectdetectionalgorithmbasedonBEVmultimodal featurefusion（BEV-MFDet）.Firstly，thealgorithmuniformlynormalizedtheimages toreducethediferenceinfeaturedistributionandappliedradiusfilteringandground fiting tothepointclouddatatoeliminatetheinfluenceofinvalidpointclouds.Atthesametime，thecolumnarnetwork integratedthespatialatentionmodule toenhanceitsabilitytolearpillrfeatures.Secondly，theBEVglobalattentionfusion moduleadaptivelyalocatedmultimodal weightstomore acuraterepresentationofscenefeatures.Experimentalresultsonthe KITTI dataset showthatthe proposedalgorithmachieves thebestoverallperformanceinmedium-diffcultydetection tasks.In particular，the detection accuracy of the small target cyclist class improves by 2.51% and 1. 75% compared to the CenterPoint andLINalgorithms withbetercomprehensive performance，respectively.Theexperimentalresults demonstrate thatthe proposedalgorithmhassuperiorperformance insmalltargetdetectionaccuracyandoveralldetectionperformance，and hashigh application value.

Key Words：BEV features；data preprocessing；global attention fusion；KITTI datasets；3D object detection

0 引言

随着人工智能技术的持续演进，自动驾驶技术已成为智能交通发展的核心方向。（剩余18597字）