回转体形状恢复与位姿估计方法

机器人抓取场景中经常包含许多具有回转结构的物体,然而由于这些物体不存在已知的三维模型,无法快速高效地确定回转体形状和位姿。针对这个问题,在现有的基于单视图回转体建模方法的基础上,将位姿估计和重建过程结合,提出回转体形状恢复与位姿估计的方法。首先利用训练好的整体嵌套的边缘检测(holistically-nested edge detection,HED)网络模型得到图像中回转体的边缘轮廓并进行细化处理,再与传统边缘检测方法相融合,提取回转体轮廓信息;其次,根据回转体成像特殊几何约束,依据单张回转体照片建立比例模型;最后,利用同一回转体不同角度的两张图像估计其尺度与位姿。通过实验表明,该方法能够通过两张同一回转体的照片,估计回转体的真实尺度与空间位姿,同时不需要点云信息,是一种鲁棒性较好的回转体形状恢复与位姿估计方法。

Research on Shape Recovery and Pose Estimation of Revolving Body

In many cases, robots need to grab objects with a revolving structure that do not have a known 3d model. In order to be able to capture these objects using existing grab point calculation schemes, their shape and pose need to be estimated. Based on the existing single-view revolving body modeling method, the pose estimation and reconstruction process are combined, and the method of shape recovery and pose estimation of the revolving body is proposed. Firstly, the edge contour of the revolving body in the image is obtained by the trained HED network model and refined, and then combined with the traditional edge detection method to extract the contour information of the revolving body. Secondly, on the basis of the special geometric constraints of the revolving body imaging, the scale model is established according to the single picture of the revolving body. Finally, two images with different angles of the same revolving body are used to estimate its scale and pose. Experiments show that the method can estimate the true scale and spatial pose of the revolving body through the images of two identical revolving body, and does not require point cloud information at the same time. It is a robust method for shape recovery and pose estimation of the revolving body.