体素化点云场景下的三维目标检测  

3D object detection in voxelized point cloud scene

作　　者：李瑞龙 吴川[1,2] 朱明 LI Rui-long;WU Chuan;ZHU Ming(Changchun Institule of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学,北京100049

出　　处：《液晶与显示》2022年第10期1355-1363,共9页Chinese Journal of Liquid Crystals and Displays

摘　　要：基于激光雷达点云数据的三维目标检测算法受制于数据量大,无法实现速度与准确率的平衡。本文提出一种改进的三维目标检测算法Pillar RCNN。首先将目标点云空间划分为体素格,使用一种基于稀疏卷积的三维主干网络将体素格逐步转化为立柱体素,三维信息量化为致密的二维信息。然后使用二维主干网络提取特征,同时将三维骨干网络中不同尺度的体素特征与二维主干网络通过多尺度体素特征聚合模块进行特征级联,通过损失函数进一步细化检测框。算法在KITTI公开数据集上进行测试,在RTX 2080Ti硬件平台上识别速度为2.48 ms。汽车、行人、自行车3种类别的检测效果同PointPillars基准算法相比较,其中自行车中等难度检测效果提升13.34%,困难难度的车检测效果提升8.85%,其他类别的检测准确率指标也有所提升,实现了速度与准确率的平衡。The 3D object detection algorithm is constrained by the large amount of point cloud data,and can not achieve the balance of real-time speed and accuracy.This paper presents an improved 3D object detection algorithm—Pillar RCNN.The algorithm firstly divides the target point cloud space into voxels,presents a 3D object detection backbone network based on sparse convolution which gradually converts voxels into column voxels,quantifies 3D information into dense 2D information,and then processes the dense 2D information through the 2D backbone network.At the same time,the voxel features of different scales in the 3D backbone network and the 2D backbone network detection results are cascaded through the multiscale voxel feature aggregation module,and the result is further refined by the loss function.The algorithm is tested on KITTI public datasets and has a recognition speed of 2.48 ms on RTX 2080Ti hardware platform.Compared with the PointPillars benchmark algorithm,the performance indicators of the three categories of car,pedestrian and bicycle are improved.The detection results of the mode of bicycle and the hard of car are improved by 13.34%and 8.85%,and the detectation accuracies of other categories are improved also,achieving a balance between speed and accuracy.

关 键 词：计算机视觉 三维目标检测 体素 稀疏卷积 特征聚合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]