A Multi-stage 6D Object Pose Estimation Method of Texture-less Objects Based on Sparse Line Features

摘要

6D pose estimation of texture-less objects is an important computer vision technique for augmented reality and vision-based robotic applications in industry. Current methods combining template and edge features suffer from large template coverage with low speed. In this paper, a new multi-stage 6D pose estimation approach with line features is proposed to solve this problem. The proposed method firstly generates a sparse template sets with the CAD model from only 4 different views. Secondly, high level geometric line features are utilized to represent the part and matched with the templates. Thirdly, sparse 2D-3D correspondences in line features and end-points are incorporated to give the coarse object pose through solving the Pn P problem. Finally, a gradient search pipeline is used to convert the object pose from coarse to fine. The experimental results reveal that the proposed approach can achieve higher accuracy and speed with a small number of templates compared to some other existing template-based methods on the Mono-6D dataset.