非结构化环境下基于占据预测的可通行性分析
作者单位:

1.上海交通大学;2.智慧地球重点实验室

基金项目:

国家自然科学基金(62273229);上海市科技创新行动计划(HCXBCY-2023-02)


WildOcc:Traversability analysis based on occupancy prediction in unstructured environment
Author:
Affiliation:

1.Shanghai Jiao Tong University;2.Key Laboratory of Digital Earth of Science

  • 摘要
  • | |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • | |
  • 文章评论
    摘要:

    为了实现在非结构化环境中的自主导航,无人车需要对地面进行可通行性分析。当前有基于激光雷达和视觉的方法对环境进行可通行性分析,但激光雷达方案受点云稀疏和成本高的限制,传统视觉方案无法有效捕捉和表达场景的三维空间状况。针对这些挑战,本文首次提出基于占据预测(Occupancy Prediction)的非结构化环境可通行性分析方法WildOcc。WildOcc从单目RGB图像提取多尺度特征,将3D占据标签投影到图像中,本文提出道路注意力机制融合信息,得到3D特征后,再经过解码器和语义分割头输出可通行区域。为了准确估计环境的三维可通行性,WildOcc使用3D占据标签进行监督;由于非结构化环境点云数据的稀疏性,本文设计了数据增强模块生成稠密占据标签(Dense Label Generate,DLG),提高监督结果的准确性本文。基于DLG模块制作了首个非结构化环境下占据预测的数据集Occ-Traversability,在该数据集上进行的综合实验表明,相对于为结构化环境设计的占据预测方法,WildOcc平均交并比(mean Intersection over Union,mIoU)提升1%;同时,DLG模块使mIoU提升0.7%,有效提升了预测的准确性。

    Abstract:

    To achieve autonomous navigation in unstructured environments, unmanned vehicles need to analyze the traversability of the terrain. Currently, methods based on LiDAR and vision are used for this analysis, but LiDAR systems are limited by sparse point clouds and high costs, and traditional vision approaches fail to effectively capture and express the three-dimensional spatial conditions of the scene. Addressing these challenges, this paper introduces for the first time a method for analyzing traversability in unstructured environments based on Occupancy Prediction, named WildOcc. WildOcc extracts multi-scale features from monocular RGB images, projects 3D occupancy labels onto the images, and introduces a road attention mechanism to query points and fuse information to obtain 3D features, which are then output as traversable areas through a decoder and semantic segmentation head. To accurately estimate the three-dimensional traversability of the environment, WildOcc uses 3D occupancy labels for supervision; due to the sparsity of point cloud data in unstructured environments, this paper designs a data enhancement module called Dense Label Generate (DLG) to produce dense occupancy labels, improving the accuracy of the supervision results. Based on the DLG module, this paper produces the first dataset usable for occupancy prediction in unstructured environments. Comprehensive experiments conducted on this dataset show that, relative to occupancy prediction methods designed for structured environments, the DLG module improves the mIoU by 0.7%, and jointly with WildOcc, enhances the mIoU by 1%, effectively increasing the prediction accuracy and robustness.

    参考文献
    相似文献
    引证文献
引用本文

刘镇,孙振,孙哲,金澄,裴凌.非结构化环境下基于占据预测的可通行性分析[J].南京信息工程大学学报,,():

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2024-07-09
  • 最后修改日期:2024-10-17
  • 录用日期:2024-10-17

地址:江苏省南京市宁六路219号    邮编:210044

联系电话:025-58731025    E-mail:nxdxb@nuist.edu.cn

南京信息工程大学学报 ® 2025 版权所有  技术支持:北京勤云科技发展有限公司