Scholarlink

Deep learning based target position estimation methods need to provide a high-precision benchmark for use as the truth value of the prediction algorithm, and there is a dearth of methods for producing spatial target datasets. To address the above problems, a method for generating target position datasets for spatial operation tasks is proposed. Firstly, the handheld camera shoots around the object to obtain video sequences. Secondly, the frame transformation was on the basis of the Aruco registration. Then, the point cloud of each frame is aligned with the point cloud of the zero frame for three-dimensional reconstruction to get the complete scene point cloud. Finally, the CAD model is registered to the scene point cloud. The transformation matrix of the object from the origin of the camera coordinate space to the zero frame of the pose is determined. In combination with the frame transformation, the transformation matrix from the target object to the camera is obtained. In this paper, models such as common satellites in space are selected as targets to validate the method. The dataset generated by the method is evaluated by testing it on the mainstream position estimation algorithms PVNet, DenseFusion, and FFB6D. The results show that the method proposed in this paper is robust and practical.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Novel Approach to Develop a Space Target Dataset for Non-cooperative Pose Estimation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Luhua Chen,
  • Dan Xiong,
  • Wei Han,
  • Chaochen Cui,
  • Yiyong Huang,
  • Yanjie Yang

摘要

Deep learning based target position estimation methods need to provide a high-precision benchmark for use as the truth value of the prediction algorithm, and there is a dearth of methods for producing spatial target datasets. To address the above problems, a method for generating target position datasets for spatial operation tasks is proposed. Firstly, the handheld camera shoots around the object to obtain video sequences. Secondly, the frame transformation was on the basis of the Aruco registration. Then, the point cloud of each frame is aligned with the point cloud of the zero frame for three-dimensional reconstruction to get the complete scene point cloud. Finally, the CAD model is registered to the scene point cloud. The transformation matrix of the object from the origin of the camera coordinate space to the zero frame of the pose is determined. In combination with the frame transformation, the transformation matrix from the target object to the camera is obtained. In this paper, models such as common satellites in space are selected as targets to validate the method. The dataset generated by the method is evaluated by testing it on the mainstream position estimation algorithms PVNet, DenseFusion, and FFB6D. The results show that the method proposed in this paper is robust and practical.