A new method for fast detection and pose estimation of texture-less industrial parts

Abstract

Estimation of the 6-Dof pose of 3d objects has been a hot research field for a long time. When robots and cameras are integrated into a system, the pose of the object can be estimated through the camera, and then the robot can be used to manipulate the object accurately. Traditional object pose estimation methods include the template-matching based method and invariant feature-based method. The method based on invariant features requires the extraction of invariant features from images with rich texture, so it is not suitable for texture-less parts, which are common in industrial applications. The template-matching method is based on edge and contour information, so it is more suitable for part detection and pose estimation of industrial applications. LINEMOD proposed by Hinterstoisser is a successful template matching method, which accelerates the template matching process through a specially designed storage structure. However, the template-based matching method generally adopts sliding window method and is very time-consuming in computation, which makes it impractical for robotic application. In this paper, we propose a new method, which combines Fully Convolutional Network (FCN) with LINEMOD algorithm. With this method, the detection and location of the object in the image can be archived quickly. Then the local image, instead of the whole image, is used for LINEMOD template matching. Experimental results show that, compared with the standard LINEMOD method, the pose estimation speed can be increased and consistent matching results can be obtained.