Accurate Pose Estimation of the Texture-Less Objects With Known CAD Models via Point Cloud Matching

Abstract

As the use of robot-based automation continues to expand in various production fields, achieving accurate pose estimation of texture-less objects has become increasingly significant and challenging. This article proposes an accurate pose estimation algorithm for texture-less objects with known CAD models by using point cloud matching. The algorithm mainly includes four parts, namely the offline data generation, global matching, local matching, and fine registration, respectively. To generate the reference point clouds required for coarse matching, the synthetic view rendering technique is utilized in the offline data generation step. To address the problem of feature matching ambiguity caused by incorrect correspondences in both the scene and reference point clouds, we introduce a similar view filter and a pose cluster mechanism in the global and local matching steps, respectively. Additionally, the genetic algorithm (GA) is utilized in the fine registration step, in which the incremental Euler angle modeling is introduced to avoid boundary cliffs and improve the searching accuracy and speed. Finally, a series of experiments on T-LESS dataset are conducted to demonstrate the effectiveness of our work. Compared with the state-of-the-art (SOTA) methods, the results show that our method improves the accuracy by at least 18% for single-input single-output pose estimation of texture-less objects.