Pose estimation for randomly piled up industrial parts based on sparse estimation

Abstract

We propose a novel method for 6D pose estimation of randomly piled up industrial parts based on sparse estimation. In this research, we consider the pose estimation problem as Least Absolute Shrinkage and Selection Operator (LASSO) that is one of the sparse optimization problems. LASSO minimizes squared residuals and L1 norm of vector which represents the solution. In this research, pose estimation is equivalent to obtain the vector which represents the selection of position and posture candidates. As our method is estimating the pose as solving a convex optimization problem, we do not need to consider the local minimum problem. Compared with other deep learning-based pose estimation algorithms, there is no need to make learning data for the network. The simulation result shows that our approach can estimate the pose of the object that has the smallest occlusion among other objects.