基于改进PointPillars的激光雷达三维目标检测算法

打开文本图片集

主题词：激光雷达 三维目标检测 体素特征稀疏卷积中图分类号：U495 文献标志码：A DOI：10.19620/j.cnki.1000-3703.20240211

LiDAR-Based 3D Object Detection Algorithm Based on Improved PointPillars

Guo Liel²，Huang Liang1， Zhao Yibing1 （1.Schoolof Mechanical Engineering，Dalian Universityof Technology，Dalian16O24;2.Ningbo InstituteofDalian University of Technology，Ningbo 315016）

【Abstract】To avoid the missng detection phenomenon caused by loss of local features due to the occlusion and truncationundercomplextraffcscenes，aLiDAR-based3DobjectdetectionalgorithmbasedonanimprovedPointPillars framework isproposed.First，thepointcloudfeatureencodingnetworkinPointPilarsisenhancedtobeterpreserveimportant structuralinformationandtherelationships withinthepointclouddata.Then，multi-scalepointcloudfeaturesareextracted andaggregated through stacked sparseconvolution modulesand submanifoldconvolution modules.Onthe bird’s-eye-view featuremapthatfusesthesemulti-scalefeatures，3Dboundingboxesofobjectsareclasifiedandregressed，improvingthe network’soperationaleficiency.Finaly，experimental testsareconductedontheKITTIdataset.Theresultsshowthat compared tothetraditionalPointPillrsalgorithm，theproposed3Dobjectdetectionmethodimproves theAveragePrecision （AP）of3Ddetectionforchallengingvehiclesamplesby6.O2percentagepoints，andincreasestheAPforsimplepedestrianand cyclistsamplesinthebird’s-eyeviewby3.35and3.18percentagepointsrespectively，efectivelyenhancing3Dobject detection accuracy.

Keywords：LiDAR，3D object detection，Pillar features，Sparse convolution 【引用格式】郭烈，黄亮，赵一兵.基于改进PointPillars 的激光雷达三维目标检测算法[J].汽车技术，2026（2）：32-38. GUOL，HUANGL，ZHAOYB.LiDAR-Based3DObject DetectionAlgorithm Basedon ImprovedPointPillars[J]. Automobile Technology，2026（2）： 32-38.

1前言

准确高效的环境感知是实现可靠自动驾驶的核心前提。（剩余11236字）