基于融合采样和图网络的三维目标检测

打开文本图片集

中图分类号：TP391.41 文献标志码：A 文章编号：1007-2683（2025）02-0042-11

Abstract：Inthe3Dtargetdetection technologybasedonpointcloud，there areproblems likehighcostof pointcloudcalculation andlargegap betwee targetscales，whichleadtolowtargetdetectioneficiency.Inresponse，thispaperproposesa3Dbject detectionalgorithmbasedonfusionsamplingandgraphetworks.Firstly，thepointcloudfusionsamplingtechnologyisintroducedto sampletheoriginalpointcloudtoeducethecomputatioalcomplexitySecondlytheK-NNalgorithisusedtoconstructtheapof thesampledpointcloud，andsub-imagesamplingisintroducedtosolvetheproblemofover-smothgraphconvolution.Finally，the featuresofgraphnodesareupdatedthroughfeatureinteraction toimprovethefeatureextractionabilityofthenetwork，thereby improvingthetargetdetectionefect.ThisstudyconductedexperimentsontheKIT3Ddataset.Comparedwith thebenchmarkmodel Point-GNN，the detection accuracy of the car target was improved by 3.89 % . In the simple scene and the medium difficulty scene，the detection accuracy of the cyclist target is increased by 6，60% and 4.36% ，respectively.

Keywords：point cloud; 3D object detection；graph neural network；fusion sampling；feature fusion

0 引言

在自动驾驶系统中，激光雷达作为环境感知设备被广泛应用。（剩余17257字）