Geo-Referenced 3D Mapping of Building Static Interiors

The concept of the geo-referenced map of the building interiors that we look forward to have. The created map of SLAM methods should get aligned with satellite maps (thus the global coordinates) and non-static objects that will change over time should be removed.

Problem Statement:

Many robotic systems rely on localization using onboard sensors such as cameras and LiDARs. When a prior 3D map of the environment is available, the robot can compare its current observations against that map to estimate its position more reliably. In practice, such prior maps are often represented as point clouds.

For robots that must operate across both indoor and outdoor environments, it is not sufficient to maintain isolated local maps only. Instead, a database of maps of multiple buildings and floors is needed, where each map is aligned with its real-world global position. This enables a robot to infer its global location even when it is operating indoors in GPS-denied areas, by localizing against a geo-referenced prior map. Another important requirement for long-term deployment is that these maps should primarily contain the static parts of the environment, excluding transient objects and short-term scene changes as much as possible.

This assignment focuses on the creation of geo-referenced 3D maps of static indoor environments for robotic localization. The student will investigate relevant methods from literature, including concepts from semantic SLAM, map alignment, and static scene extraction. Basic knowledge of robotics software frameworks such as ROS, as well as familiarity with 3D sensing and mapping, will be beneficial. The work will involve both analysis of existing methods and practical experimentation with recorded data collected by our robotic platforms.

Tasks:

·      Study literature on 3D mapping, semantic SLAM, and methods for separating static and dynamic elements in indoor environments.

·      Investigate strategies for geo-referencing indoor point cloud maps so that multiple buildings and floors can be integrated into a globally consistent map database.

·      Work with recorded real-world sensor data from our robots, including camera and/or LiDAR data.

·      Benchmark relevant mapping and map-processing algorithms with respect to map quality, static-scene consistency, and suitability for localization.

·      Develop the necessary software tools for integration of the selected methods into a robotics workflow, preferably in ROS-based environments.

Work:

30% Theory, 50% Benchmarking and Development, 20% Writing

Contact:

Le Viet Duc – Pervasive Systems, EEMCS, University of Twente

Hojat Mirtajadini, SMART Mechatronics and Robotics Research Group, Saxion University of Applied Sciences