作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (5): 262-268. doi: 10.19678/j.issn.1000-3428.0065910

• 图形图像处理 • 上一篇    下一篇

火星表面影像分割数据集及性能评估

吕炜琨1, 魏琳慧1, 郑殿1, 刘雨1,2, 望育梅1   

  1. 1. 北京邮电大学 人工智能学院, 北京 100876;
    2. 鹏城实验室, 广东 深圳 518000
  • 收稿日期:2022-10-04 修回日期:2023-02-09 发布日期:2023-03-22
  • 作者简介:吕炜琨(2000-),男,硕士研究生,主研方向为图像处理、深度学习;魏琳慧,博士研究生;郑殿,硕士研究生;刘雨(通信作者),副教授、博士、博士生导师;望育梅,副教授、博士。
  • 基金资助:
    国家重点研发计划(2019YFB1803103)。

Martian Surface Image Segmentation Dataset and Performance Evaluation

Lü Weikun1, WEI Linhui1, ZHENG Dian1, LIU Yu1,2, WANG Yumei1   

  1. 1. College of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876, China;
    2. Pengcheng Laboratory, Shenzhen 518000, Guangdong, China
  • Received:2022-10-04 Revised:2023-02-09 Published:2023-03-22

摘要: 随着深空探测技术的不断进步,中国火星探测任务天问一号搭载的“祝融号”火星车成功着陆火星,并开展火星表面影像采集工作。火星表面的主要障碍物是岩石,这些形状、大小不同的岩石将会影响火星车在科学探索任务中的路径规划。准确的岩石分割对于火星车避障和路径规划具有重要意义。由于目前对于天问一号所拍摄的图像数据集应用较少,且多数火星图像的分割数据集处于非公开状态,在一定程度上阻碍了火星岩石分割相关算法的研究。为此,面向“祝融号”火星车上搭载的导航地形相机所拍摄的火星表面影像,通过人工进行图像筛选和标记,建立一个火星表面影像分割数据集TWMARS。该数据集共包含336张火星图像,将其按比例随机划分为训练集、验证集和测试集,在现有的语义分割算法上进行数据集性能评估实验,结果表明,在MobileNet、ResNet等系列卷积神经网络模型上运行TWMARS数据集,平均像素精度和平均交并比最高分别可达83.0%和77.4%,在经典语义分割算法上获得较高的准确率,可应用于火星岩石检测算法的相关研究。

关键词: 天问一号任务, 火星车, 导航地形相机, 火星表面影像, 分割数据集

Abstract: As part of the continuous progress being made in deep space exploration technology,the “Zhurong Mars ”rover carried by China's first Mars exploration mission,Tianwen-1,has successfully landed on Mars and collects images of the Martian surface.Rocks are the main obstacle on the surface of Mars,with rocks of different shapes and sizes having an impact on the path planning of the rover during scientific exploration missions. Accurate rock segmentation is of great significance for obstacle avoidance and path planning.Because there are few applications for the image datasets captured by Tianwen-1,the segmentation datasets from most Mars images are restricted to the non-public state,essentially hindering the research into the algorithms related to Mars rock segmentation.Therefore,the images of the Martian surface,captured by the Navigation and Terrain Cameras(NaTeCam) mounted on the “Zhurong” Mars rover,were manually screened and annotated to establish the Martian surface image segmentation dataset,TWMARS.The dataset contained 336 Martian surface images,which were randomly divided into the training,validation,and test sets according to their required proportions.Performance evaluation tests were conducted on the existing semantic segmentation algorithms of the dataset.The TWMARS data set were run using the MobileNet and ResNet Convolutional Neural Network(CNN) models achieved the highest mean pixel accuracy and mean intersection over union values of 83.0% and 77.4%,respectively.The experimental results show that the classical semantic segmentation algorithms of the dataset achieved a high accuracy and can be applied further to the research regarding the Mars rock detection algorithm.

Key words: Tianwen-1 mission, Mars rover, Navigation and Terrain Cameras(NaTeCam), Martian surface image, segmentation dataset

中图分类号: