High-Speed Pose Tracking for Robotic Grasping Using an RGB Camera and Hybrid Vision Techniques

Object pose tracking is a crucial part in automated robotic manipulation, requiring reliably updating the pose estimate at high frequencies so to enable precise high-bandwidth control in real-time applications. This thesis presents a pipeline for model-based pose-tracking with an RGB camera, implementing deep-learning-based algorithms for object segmentation and initial pose estimation to guarantee flexibility, while classical computer vision methods are used to iteratively update the pose in each image frame. The pipeline achieves an average update rate of 50 Hz while demonstrating good tracking performance for objects in various scenarios, including a robotic grasping experiment. However, challenges remain in accurately estimating the orientation of symmetrical objects and handling objects with similar colours to the background or highly textured surfaces.