This thesis A Multi-task Robot-vision System Based on Pose Estimation Algorithms presents a multi-task hand-eye system for multiple tasks such as workpiece loading and unloading, bin picking, workpiece 3d reconstruction, workpiece printing inspection and so on. An accuracy and high-resolution 3d camera is mounted on the end-effector of the industry robot arm which forms a hand-eye system. We established a system that integrates all necessary components including hand-eye calibration, robot tool calibration, robot communication, multi-view point cloud fusion, workpiece pose estimation and grasp pose selection.

Because of the reflective nature of metal workpieces and the difficulty of pose estimation problems, we have to fuse multiple views together. We have implemented the TSDF...[ Read more ]

This thesis A Multi-task Robot-vision System Based on Pose Estimation Algorithms presents a multi-task hand-eye system for multiple tasks such as workpiece loading and unloading, bin picking, workpiece 3d reconstruction, workpiece printing inspection and so on. An accuracy and high-resolution 3d camera is mounted on the end-effector of the industry robot arm which forms a hand-eye system. We established a system that integrates all necessary components including hand-eye calibration, robot tool calibration, robot communication, multi-view point cloud fusion, workpiece pose estimation and grasp pose selection.

Because of the reflective nature of metal workpieces and the difficulty of pose estimation problems, we have to fuse multiple views together. We have implemented the TSDF fusion algorithm and accelerate it with CUDA.

We have developed a novel Hough voting based pose estimation algorithm which obtained results close to the stat of the art. We combined Hough voting strategy with 3D learning networks. In addition, the algorithm takes multi-view information as input and output poses of potential workpieces. Moreover, we have implemented and improved the point pair feature based algorithm in several ways to further improve the performance of the algorithm.