POSE ESTIMATION OF DAILY CONTAINERS FOR A LIFE-SUPPORT ROBOT

摘要

Caring for people who are elderly is now a global challenge given the short-age of nursing and healthcare providers. This makes life-support robots that could assistpeople to live independently highly significant. A frequent task during life support is thefetching of daily containers, which requires accurate sia-dimensional pose estimation. Inthis paper, we develop a pipeline that is capable of providing such estimation. Firsttransfer learning is used to retrain an object-detection model based on a convolutionalneural netuork to produce accurate rectangular masks. After extracting object cloudsbased on these masks, the iterative closest point algorithm is used to perform point-cloudregistration, resulting finally in pose estimation. Several approaches are introduced to increase the registration accuracy and stability by providing adequate initial alignment andsuitable model clouds considering both self-occlusions and partial occlusions. Through ea-periments involing actual daily scenarios, the effectiveness and accuracy of the proposedpipeline are verified.