Fast and precise 6D pose estimation of textureless objects using the point cloud and gray image

摘要

Pose estimation for textureless objects is a challenging task in robotics, due to the scanty information of surfaces. In this paper, we design a vision system for fast and precise position and orientation measurement of textureless objects with a depth camera and a CCD camera. The corresponding process includes two parts: object segmentation in the point cloud and pose measurement in the gray image. Considering the relation between the object and its fixed panel, we first extract the panel in the point cloud by combining a random sample consensus algorithm with local surface normal. We then coarsely segment the possible area of the object based on an oriented bounding box. Finally, we transform the point cloud coordinates into the image coordinate system, and measure the precise pose of the object with a view-based matching method. Two types of cameras are brought together to make their respective advantages play well. The downscale method and coarse-to-fine strategy are utilized sufficiently to increase efficiency. Experiments show that our vision system achieves high pose measurement precision and enough efficiency. The average error is less than 2 mm for x and y, less than 4 mm for z, and 1° in orientation, meeting the requirements for our robotic grasping task.