Real-time 3D Semantic Mapping based on Keyframes and Octomap for Autonomous Cobot

This paper presents an approach to develop a semantic mapping framework based on RTABMap as a SLAM algorithm and Octomap as the data structure to represent the 3D map of indoor environments. The reason behind choosing Octomap as the main visual representation comes from its ability to deal with environments where the objects are dynamic while providing enough details of the environment in case of teleoperation. Adding semantics to Octomap does not only allow autonomous navigation, but could also be used as an assisting tool for the tele-operator for safer navigation. In this framework, the semantics are integrated inside the Octomap using Bayesian fusion and the integrated confidence of the semantic class of each voxel is stored inside the voxel itself. This allows us to take the uncertainty of the semantic object generator in consideration when classifying the voxels of the Octomap, hence this framework can also be used in grasping and manipulation scenarios where the semantics of the scene are not taken for granted and a confidence measure is used before a decision is taken.