Pose Estimation of Texture-Less Targets for Unconstrained Grasping

With the advent of Industry 4.0, the demand for estimating target pose keeps increasing. However, the accuracy of the existing pose estimation algorithms for texture-less targets is still poor. Traditional methods require approximately accurate initial pose or else they are easy to fall into a local optimum while the deep learning methods are limited in unconstrained environments where the unpredictable data can not be captured ahead for model training. Therefore, the paper proposes an innovative method which can cover the shortage of these two classes of methods. In our method, a multi-task model which can predict the pose of target and simultaneously obtain the edge map is designed. Then, the predicted pose and edge map are transferred to pose optimization module which is implemented based on edge matching. In addition, considering the lack of the pose datasets for texture-less objects, we design an effective pose dataset generation method based on 3D reconstruction. At last, the proposed system is tested on the public dataset and the rendered dataset. Experimental results demonstrate that the proposed algorithm is more accurate compared with the state-of-the-art methods.