Object recognition and pose estimation for modular manipulation system: overview and initial results

Object detection and pose estimation is a fundamental functionality among robotic perception for manipulation. Applying robots to diverse tasks requires a robust perception skill. In this manuscript, we introduce an overview of our object recognition and pose estimation process and its our initial results. Our approach follows the previous approaches using local feature extraction and match. As a training stage, synthetic dataset is generated with its 2D-3D information. Local features is extracted and its 2D-3D information are stored in the dataset. As a test stage, the background area is removed and blobs which might include object candidates are extracted. Then, the local features are extracted and matched with the features stored in the database and the correspondences are found. Based on the correspondences, object instance and pose information is estimated by solving Perspective-n-Point problem. To validate our approach, we trained the system with synthetic images and tested it with real images for object recognition and detection and with synthetic images for object pose estimation.