An intelligent robotic vision system with environment perception

Ever since the dawn of computer vision[1, 2], 3D environment reconstruction and object 6D pose estimation have been a core problem. This thesis attempts to develop a novel 3D intelligent robotic vision system integrating environment reconstruction and object detection techniques to solve practical problems. Chapter 2 reviews current state-of-the art of 3D vision techniques from environment reconstruction and 6D pose estimation.In Chapter 3 a novel environment reconstruction system is proposed by using coloured point clouds. The evaluation experiment indicates that the proposed algorithm 2 is effective for small-scale and large scale and textureless scenes. Chapter 4 presents Image-6D (that is section 4.2), a learning-based object pose estimation algorithm from a single RGB image. Contour-alignment is introduced as an efficient algorithm for pose refinement in an RGB image. This new method is evaluated on two widely used benchmark image data bases, LINEMOD and Occlusion-LINEMOD. Experiments show that the proposed method surpasses other state-of-the-art RGB based prediction approaches. Chapter 5 describes Point-6D (defined in section 5.2), a novel 6D pose estimation method using coloured point clouds as input. The performance of this new method is demonstrated on LineMOD [3] and YCB-Video [4] dataset. Chapter 6 summarizes contributions and discusses potential future research directions. In addition, we presents an intelligent 3D robotic vision system deployed in a simulated/laboratory nuclear waste disposal scenario in Appendices B. To verify the results, a simulated nuclear waste handling experiment has been successfully completed via the proposed robotic system.