汽车工程 ›› 2022, Vol. 44 ›› Issue (11): 1656-1664.doi: 10.19562/j.chinasae.qcgc.2022.11.004

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

基于自注意力机制的自动驾驶场景点云语义分割方法

王大方1,尚海1,曹江1(),王涛2(),夏祥腾1,韩雨霖1   

  1. 1.哈尔滨工业大学(威海)汽车工程学院,威海  264200
    2.陆军装甲兵学院,北京  100072
  • 收稿日期:2022-05-08 修回日期:2022-06-17 出版日期:2022-11-25 发布日期:2022-11-19
  • 通讯作者: 曹江,王涛 E-mail:1964611621@qq.com;3387340132@qq.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 Point Cloud in Automatic Driving Scene Based on Self-attention Mechanism

Dafang Wang1,Hai Shang1,Jiang Cao1(),Tao Wang2(),Xiangteng Xia1,Yulin Han1   

  1. 1.School of Automotive Engineering,Harbin Institute of Technology,Weihai  264200
    2.Army Academy of Armored Forces,Beijing 100072
  • Received:2022-05-08 Revised:2022-06-17 Online:2022-11-25 Published:2022-11-19
  • Contact: Jiang Cao,Tao Wang E-mail:1964611621@qq.com;3387340132@qq.com

摘要:

对车载激光雷达场景点云进行语义分割是自动驾驶环境感知环节的基础性工作。针对现有处理大规模自动驾驶场景点云方法对局部特征提取能力不足和难以捕捉全局上下文信息的问题,本文基于自注意力机制设计了局部和全局自注意力编码器,并搭建了特征聚合模块进行特征提取。实验结果表明,与同样采用局部特征聚合的网络RandLA-Net相比,在SemanticKITTI数据集上本文的方法可将平均交并比提升5.7个百分点,局部自注意力编码器的加入也使车辆和行人等小目标的分割精度提高2个百分点以上。

关键词: 语义分割, 大规模点云, 自注意力机制

Abstract:

Semantic segmentation of vehicle lidar scene point cloud is the basic work of automatic driving environment perception. In view of the insufficient ability to extract local features and difficulty to capture the global context information of the existing processing method of point cloud in large-scale automatic driving scene, the local and global self-attention encoders are designed based on the self-attention mechanism and the feature aggregation module is built for feature extraction. The experimental results show that compared with RandLA-Net, also adopting local feature aggregation, the method proposed can increase the MIoU by 5.7 percentage points on SemanticKITTI dataset, and adding local self-attention encoder also raises the segmentation accuracy of small targets such as vehicles and pedestrians by more than 2 percentage points.

Key words: semantic segmentation, large-scale point cloud, self-attention mechanism