基于局部信息融合的点云3D目标检测算法

张林杰,柴志雷,王宁

Point cloud 3D object detection algorithm based on local information fusion

Linjie ZHANG,Zhilei CHAI,Ning WANG

表 1 KITTI测试数据集上不同算法的检测结果对比

Tab.1 Comparison of detection result from different algorithm on KITTI test dataset

方法 模态 AP3D/% APBEV/% 简单 中等 困难 mAP 简单 中等 困难 mAP Point-GNN[16] L 88.33 79.47 72.29 80.03 93.11 89.17 83.90 88.73 3DSSD[18] L 88.36 79.57 74.55 80.83 92.66 89.02 85.86 89.18 PV-RCNN[30] L 90.25 81.43 76.82 82.83 94.98 90.65 86.14 90.60 Voxel-RCNN[9] L 90.90 81.62 77.06 83.19 94.85 88.83 86.13 89.94 CT3D[8] L 87.83 81.77 77.16 82.25 92.36 88.83 84.07 88.42 Pyramid-PV[28] L 88.39 82.08 77.49 82.65 92.19 88.84 86.21 89.08 VoTr[21] L 89.90 82.09 79.14 83.71 94.03 90.34 86.14 90.17 SPG[22] L 90.50 82.13 78.90 83.84 94.33 88.70 85.98 89.67 VoxSet[27] L 88.53 82.06 77.46 82.68 — — — — PDV[31] L 90.43 81.86 77.36 83.22 94.56 90.48 86.23 90.42 VFF[32] L+I 89.50 82.09 79.29 83.62 — — — — PG-RCNN[23] I 89.38 82.13 77.33 82.88 93.39 89.46 86.54 89.80 PVT-SSD[24] I 90.65 82.29 76.85 83.26 95.23 91.63 86.43 91.10 DVF-PV[25] L+I 90.99 82.40 77.37 83.58 — — — — 本文方法 L 91.60 82.53 77.83 83.99 95.59 91.37 86.72 91.23