SEMANTIC PERCEPTION IN DAILY INDOOR ENVIRONMENTS

Abstract

As an essential component for autonomous mobile manipulation in daily indoor environments, the robotic semantic perception system is in charge of perceiving the environment and generating semantic perception results, answering the questions of “Where and what are the objects?”, “What is the place like?”, “Where could the robotic gripper be placed to grasp the objects?”, etc. This thesis focuses on the problem of semantic perception in daily indoor environments. Three fundamental semantic perception tasks are identified, 2D object detection and scene recognition, tabletop robotic grasp detection, and amodal 3D object detection. We propose several semantic perception algorithms using Convolutional Neural Networks. We evaluate our algorithms on public datasets and data collected in the real world. Experiments show that our algorithms achieve outstanding performances compared to state-of-the-art methods.