迭代伪点云生成的3D目标检测
打开文本图片集
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
近年来随着自动驾驶技术的快速发展，人们对车辆感知和理解周围环境的要求不断提高，3D目标检测技术受到了极大的关注。（剩余14348字）