迭代偽點云生成的3D目標(biāo)檢測
打開文本圖片集
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
近年來隨著自動駕駛技術(shù)的快速發(fā)展，人們對車輛感知和理解周圍環(huán)境的要求不斷提高，3D目標(biāo)檢測技術(shù)受到了極大的關(guān)注。（剩余14348字）