基于毫米波雷达和相机融合的3D目标检测研究

打开文本图片集

中图分类号：TN911.73-34 文献标识码：A 文章编号：1004-373X（2026）07-0120-07

引用格式：，，，等.基于毫米波雷达和相机融合的3D目标检测研究[J].现代电子技术，2026，49（7）：120-126

Abstract：Duetothesparsityofthepointcloudsintheprocessofsensorfusion，thereisalackofgeometricinforationin thelowreflectionareaofsmallobjects，whichleadstothedificultyinaligningthefeaturesoftheimageandradarpointclouds， affectingtheefectivefusionofradarandcamerainformation.Inviewof theabove，a3Dobjectdetectionalgorithmbasedonthe fusionofmillmeterwavedarandcameraisproposed.Thealgorithm，namedRE-BEVDepth，isimprovedandoptimizedintwo aspects.Ontheonehand，themodelPointPillasisused toobtainthefeatureinformationofmilimeterwaveradarpointclouds andmapittothe pseudo image.Thepseudo image featuresare extractedandfused withtheimage features obtainedbymodel BEVDepth inBEVspace.Ontheotherhand，theBackbonenetworkissimplified，high-levelfeaturesareextractedfromthe pseudoimagesgeneratedbythemilimeterwaveradar，soastoobtainthefeaturesofthebird's-eyeview（BEV）.Experimental results onthedatasetnuScenes showthatthe mean average precision （mAP）ofthe proposedalgorithm is6.99% higherthan that ofBEVDepth，and the model reasoning duration is reduced by 6.14ms ，which proves that the algorithm hasmore accurate perceptionability，andcanfurther meetthedetectionrequirementsofautomaticdriving technology inenvironmentperception.

Keywords：3Dobjectdetection; multi-sensor fusion;milimeter-waveradar;BEV;automaticdriving; neuralnetwork

0 引言

近年来，自动驾驶技术进展显著，通过智能化感知和决策实现了车辆在无人条件下的正常行驶。（剩余10765字）