Combining Global and Local 6DoF Pose Estimation for Real-World Applications

Abstract

In this thesis, we tightly integrate object detection, pose estimation and tracking into a single, effective pipeline. For object pose detection on resource constrained robotic hardware, we train a lightweight SSDlite detector and Augmented AutoEncoders. To avoid manual labeling, we only train on photorealistic data generated using BlenderProc. We investigate photo-metric augmentations to reduce the sim2real gap. Furthermore, we develop techniques for tracking loss detection based on features and metrics extracted from an existing tracker, as well as the trained bounding box detector and pose estimator. The goal is thereby to automatically alternate between accurate and fast object tracking and (re-)initialization with global object pose detection. For tracking loss detection, different statistical and learning-based methods are evaluated on the YCB-Video dataset. Finally, we test our pipeline on the anthropomorphic robot David for 6-DoF pose estimation and tracking of IKEA dishes.