Real-time pose estimation of rigid objects using RGB-D imagery

摘要: Using full scale (480×640) RGB-D imagery, we here present an approach for tracking 6d pose of rigid objects at runtime frequency up to 15fps. This is useful robotic perception systems efficiently track object's during camera movements in tabletop manipulation tasks with high detection rate and real-time performance. Specifically, appearance-based feature correspondences are used initial object detection. We make use Oriented Brief (ORB) key-points perform fast segmentation candidates the 3d point cloud. The task estimation handled Cartesian space by finding interest window around segmented geometry operations. later extraction subsequent frames speed process. also allows efficient scenes where there significantly large false matches between due scene clutter. Our was tested using dataset comprising from video sequences tabletops multiple household environments. Experiments show that our capable performing followed higher frame rates compared existing techniques.