汽车工程 ›› 2022, Vol. 44 ›› Issue (1): 26-35.doi: 10.19562/j.chinasae.qcgc.2022.01.004

所属专题: 智能网联汽车技术专题-感知&HMI&测评2022年

• • 上一篇    下一篇

基于稀疏卷积神经网络的车载激光雷达点云语义分割方法

夏祥腾,王大方(),曹江,赵刚,张京明   

  1. 哈尔滨工业大学(威海)汽车工程学院,威海  264200
  • 收稿日期:2021-10-11 修回日期:2021-10-25 出版日期:2022-01-25 发布日期:2022-01-21
  • 通讯作者: 王大方 E-mail:13863009863@163.com
  • 基金资助:
    哈尔滨工业大学重大科研项目培育计划(ZDXMPY20180109)

Semantic Segmentation Method of On-board Lidar Point Cloud Based on Sparse Convolutional Neural Network

Xiangteng Xia,Dafang Wang(),Jiang Cao,Gang Zhao,Jingming Zhang   

  1. School of Automotive Engineering,Harbin Institute of Technology,Weihai  264200
  • Received:2021-10-11 Revised:2021-10-25 Online:2022-01-25 Published:2022-01-21
  • Contact: Dafang Wang E-mail:13863009863@163.com

摘要:

对车载激光雷达扫描得到的点云进行语义分割是保证行车安全、加强驾驶员对周边环境理解的重要手段之一。因为内存限制和大规模点云场景更加稀疏的特点,将传统神经网络的方法直接沿用到车载激光雷达扫描得到的点云场景中的效果不佳。本文中针对大规模点云的稀疏性,利用稀疏卷积神经网络对体素化点云进行特征提取。考虑到逐点处理分支抑制点云数据的密度不一致性导致的信息损失,另外设计了3D-CA和3D-SA模块,使稀疏卷积神经网络更好地提取特征。实验结果表明,与传统卷积神经网络的方法和将点云投影到平面的方法相比,使用稀疏卷积神经网络对大规模点云进行语义分割,可将平均交并比提升4.1%和3.4%,证明了该方法的有效性。

关键词: 无人驾驶, 点云, 语义分割, 稀疏卷积神经网络

Abstract:

The semantic segmentation of point cloud obtained by on-board lidar scanning is one of the important means for ensuring driving safety and enhancing the driver's understanding of surrounding environment. Due to memory limitation and the sparse characteristics of large-scale point cloud scenes, directly continue to apply the traditional neural network approach to the large-scale point cloud scenario does not work well. Therefore, taking the advantage of the sparseness of large-scale point cloud, sparse convolutional neural network (sparse CNN) is used to extract the characteristics of voxel cloud in this paper. With consideration of the information loss cause by density inconsistence in point-wise processing sub-branch suppressed point cloud data, additional 3D-CA and 3D-SA modules are designed to improve the characteristics extraction of sparse CNN. The results of experiment show that compared with traditional convolutional neural network method and point cloud projection on plane method, the sparse CNN method for the semantic segmentation of large scale point cloud can have 4.1% and 3.4% higher IOU respectively, demonstrating the effectiveness of the method adopted.

Key words: autonomous driving, point cloud, semantic segmentation, sparse convolution neural network