迭代偽點(diǎn)云生成的3D目標(biāo)檢測(cè)

打開(kāi)文本圖片集

3D object detection based on iterative pseudo point cloud generation

Sun Lihui?，Wang Chuyao （SchoolofManagementScience&IforationEnginering，HebeiUniersityfconomics&Businss，ijzangO5OChina）

Abstract：3Dobject detection iscrucial forautonomous driving.However，incomplex scenarios，LiDAR oftenstruggles to capture complete point-clouddatadue todistance andocclusion，afectingdetection accuracy.To addressthis，the paperpro poseda3Dobject detectionmethodbasedoniterativepseudo-point-cloudgeneration（IG-RCNN）.Firstly，itintroduceda channel sparsepartialconvolution（CSPConv）module inthe3Dvoxel backbone toreduce channel redundancyand fuse semanticinformationfrom diferentreceptivefields，enhancing feature fusion.Secondly，iterativerefinementgeneratedhighqualitypseudo-pointclouds，providing efectiveguidanceforthesuggestionboxandimprovingdetectionacuracy.Experiments on the KITTI dataset show that the algorithm outperforms PV-RCNN，with a 3.89% and 2. 73% accuracy improvement for pedestrians andcyclists，respectively，under harddificulty.Thisdemonstrates thealgorithm’ssuperiorityinprocesingsparse point clouddata，especiallyindetectingsmallojects likepedestrians and cyists，shows strongerrobustnessandaccuracy

Key words：autonomous driving；driver asistance system；3D object detection；pseudo-point cloud generation

0 引言

近年來(lái)隨著自動(dòng)駕駛技術(shù)的快速發(fā)展，人們對(duì)車(chē)輛感知和理解周?chē)h(huán)境的要求不斷提高，3D目標(biāo)檢測(cè)技術(shù)受到了極大的關(guān)注。（剩余14348字）