基于卷积神经网络的激光雷达点云目标分割  

Convolutional Neural Nets for Real-time Road-object Segmentation from 3D LiDAR Point Cloud

作　　者：张青[1] 黄影平[1] ZHANG Qing;HUANG Yingping(University of Shanghai for Science and Technology University,Shanghai 200093,China)

机构地区：[1]上海理工大学,上海200093

出　　处：《通信技术》2021年第7期1634-1640,共7页Communications Technology

摘　　要：对激光雷达点云进行道路对象的语义分割,尤其是对感兴趣实例(如汽车、行人和自行车)的检测与归类。将此问题明确表达为逐点分类问题,并以卷积神经网络(Convolutional Neural Nets,CNN)为基准网络设计网络结构,对预处理过的点云数据进行语义分割。具体地,CNN将转化后的雷达点云数据作为输入,直接输出逐点标号的预测图,再通过条件随机场(Conditional Random Field,CRF)对预测图进行完善。CNN模型在KITTI数据集产生的LiDAR点云上训练,逐点分割标签源自KITTI产生的3D边界框。实验表明,设计的网络结构基本满足自动驾驶所需要的高精确度和快处理速度(每帧13.8 ms左右)。This paper performs semantic segmentation of road objects from 3 D LiDAR point cloud,especially the detection and categorization of interesting instances(such as cars,pedestrians,and bicycles).This problem is clearly expressed as a point-wise classification problem,and an end-to-end pipeline based on convolutional neural networks(CNN)is proposed:the CNN takes a transformed LiDAR point cloud as input and directly outputs a point-wise label map,which is then refined by a conditional random field(CRF),implemented as a recurrent layer.The CNN model is trained on LiDAR point clouds from the KITTI dataset,and point-wise segmentation labels are derived from 3 D bounding boxes from KITTI.The experiments results indicate that the pipeline achieves high accuracy with astonishingly fast and stable runtime(about 13.5 ms per frame),highly desirable for autonomous driving applications.

关 键 词：自动驾驶 3D LiDAR点云 点云转换 卷积神经网络 条件随机场 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]