3D-Multi-Layer-Multi-Representation-Maps for Short- and Long-term Mapping and Navigation

Robots are constantly becoming more integral in the day-to-day lives of humanity. To do this, they have to accomplish tasks autonomously in dynamically changing environments. Dynamic objects often need to be handled differently than static objects because they cause changes in the environment. To solve the problem, we propose a 3D-Multi-Layer-Multi-Representation map. The overall map consists of multiple layers with custom semantics and custom representation types. A static layer models the static environment using an octomap. A second layer models generic dynamic objects as bounding boxes. Semantic segmentation is used to decide which measurement belongs to a dynamic object. These are all objects of beforehand defined classes. This allows customized update strategies for both types of objects. The experiments show that this increases the accuracy and efficiency of the overall map, as well as the individual layers. A third layer, the human layer that stores the poses of all persons, is added to the map. This allows to precisely see what a human is doing in the exact moment. Using different representations types, the overall map does not only have a higher accuracy and efficiency, but also provides more in-depth knowledge of the scene.