Pose Estimation Method Combining the PnP Algorithm and Contour Depth Extraction for Peg-in-Hole Assembly

Abstract

There is ambiguity in the hole pose estimation during vision-guided robotic arm peg-in-hole tasks. This paper proposes a pose an estimation method combining the PnP algorithm and contour depth extraction to address the ambiguity problem. Yolov5 is employed to segment the unstructured scene. After Canny edge detection and contour extraction are performed on the segmented images, the circumference of the circular hole is fitted with the least squares method to obtain an oval shape. The hole pose is estimated through two stages of visual recognition. In the first recognition stage, the camera's optical axis is aligned with the circle's center to determine its position precisely. In the second recognition stage, the PNP algorithm is employed with the extracted contours to ascertain the hole's pose. To resolve the ambiguity in pose estimation when the assembly hole is tilted, the PNP algorithm is initially utilized to estimate the pose of the assembly hole. Subsequently, based on the obtained contour information, extract the depth values of the two parts of the ellipse separated by the major axis. By comparing these values, determine the rotation direction around the y-axis, thereby obtaining the unambiguous pose of the assembly hole. The experiment results show that the average position error of the circular hole is within 0.8mm, and the average pose error is within 4°. This method can effectively solve the ambiguity problem in the pose estimation of Peg-in-Hole assembly.