3D Pose Recognition of Small Special-Shaped Sheet Metal with Multi-Objective Overlapping

This paper addresses the challenging task of determining position and posture small-scale thin metal parts with multi-objective overlapping. To tackle this problem, we propose a method that utilizes instance segmentation three-dimensional (3D) point cloud for recognizing special-shaped parts. We investigate process obtaining single target by aligning mask depth map. Additionally, explore pose estimation involves registering model cloud, designing registration algorithm combines sample consensus initial alignment (SAC-IA) coarse iterative closest (ICP) fine registration. The experimental results demonstrate effectiveness our approach. average accuracy models, utilizing ResNet50 + FPN ResNet101 as backbone networks, exceeds 97%. time consumption is reduced 50%. Furthermore, algorithm, which SAC-IA ICP, achieves lower while satisfying requirements manufacturing new energy batteries.