Robots that are expected to navigate efficiently through real-world environments need maps to localize themself and plan actions. These maps just contain purely geometric information as the robot needs a semantic understanding of its surroundings to fulfill their tasks. In this project, we address the problem of object-based mapping for mobile robots in indoor environments. The goal of the project is to realize a mapping system where an occupancy representation of the scene is enhanced with object level semantics. This high-level understanding of the scene should be suitable for other robots to localize and plan actions.
A gazebo simulation of an indoor environment is used. Through this environment we move a robot which collects RGBD and Lidar data.
segmentation by multiplying the past probabilities for each label with new probabilities for the label in the same cell
segmentation by adding the label probabilities for the same cell and finding maximum probability after softmax in each cell of occupancy grid
