Object Detection for Texture-less Tubular Joints using Hierarchical CAD Tree

Object detection for large tubular joints is one of the major challenges for robotic welding. These tubular joints are texture-less objects with a shiny and strongly reflective surface. Additionally, they are attached with large fixtures resulting in dense self-occlusion. Due to this, it is difficult to plug-and-play existing 3D object recognition pipelines. This paper presents a hierarchical CAD tree (HCT), generated using a 3D CAD model at the root node, partial views as successive layer nodes and segmented-partial views, as leaf nodes of the tree. This hierarchical approach stores the assembly information of the leaf nodes that is used for our hypothesis verification pipeline. The complete method is tested using a RGB-D sensor mounted on a robotic manipulator on a gantry, to detect tubular joints in a shipyard environment. The framework demonstrates that the proposed approach can detect horizontal and diagonal configurations of tubular joints despite high levels of occlusion.