Semantic Mapping for Indoor Robotics

Abstract

Advancements in autonomous robotics research have reached a stage where mobile robots are being deployed in social environments such as offices, malls, and households. Mobile robots deployed in such crowded environments often need to interact with objects to perform a task or to navigate through the environment to reach their destination. It is imperative that the robot has knowledge about the semantics and the complete geometry of the object it needs to interact with, because interaction with incomplete information can cause safety risks to the object or the surroundings. The goal of this thesis was to develop a method that generates semantic maps with complete geometric information of objects in the environment using color and depth sensors. To this end, a pipeline was implemented to construct a mesh representation of the environment with partial object representations in the mesh replaced with similar synthetic models. The pipeline also explores the viability of a deep learning based approach to complete the missing regions in the object representations. Tests were conducted to evaluate the pipeline on an office environment simulated in Gazebo and on a real environment using a Care-O-bot robot. The pipeline performs well when the pace of the robot navigating the environment is low. The comparison between the deep learning based approach and the synthetic model matching approach favors the latter for better quality output. Furthermore, the analysis of results indicated that the 2D image segmentation method used limits the performance of the pipeline as a whole. Future directions to expand the work can include expansion of the database to accommodate multiple object classes, using an alternate method for image segmentation, and exploring the possibility of applying the output generated by the pipeline for 'Interactive Navigation' research.