Pose Estimation of a Simple-Shaped Object Based on PoseClass Using RGBD Camera

Abstract

The problem of pose estimation of a simple-shaped object using an RGBD camera is addressed with the purpose of developing a robot capable of arranging goods. The demand for robots capable of arranging goods in retail stores is high. However, the goods are usually simple in shape such as a rectangular box or triangular prism, and it is difficult to estimate the pose using conventional methods based on shape features, without a given rough pose as an initial value. In this study, a new concept called PoseClass is proposed, in which the object surface placed on the shelf is treated as a class, and a Deep Neural Network (DNN) is developed, which while estimating the PoseClass also outputs the pose. The developed method is 3.8 times more accurate than previous DNN-based methods.