mit-acl/meridian

Meridian is a research software package from MIT that solves a specific navigation problem for ground robots: figuring out where you are when you only have an overhead aerial photo to compare against. A robot driving through an outdoor environment can use Meridian to determine its precise location on a map derived from aerial imagery, with accuracy down to roughly one meter, and without needing a known starting position or any location-specific tuning beforehand. The approach works by extracting simple geometric shapes, specifically points and lines, from both the aerial view and the robot's own camera view. The system then finds matches between these shapes across the two very different perspectives. To handle the fact that a sky-down view and a ground-level view look quite different, Meridian builds up a series of pose measurements over time and uses a method called pose graph optimization to find the most consistent set of location estimates. The code is written primarily in Python, with computationally heavy parts handled in C++ or using a graphics card through PyTorch. A connector for ROS2, the standard software framework used in robotics research, is listed as coming soon. The repository includes an install script that sets up the environment, along with a script that downloads about 8 gigabytes of demo data so you can test the full pipeline immediately after installation. This is a research project accompanied by a paper on arXiv from a team at MIT and collaborators. The full dataset used in the research, called the Camp Dataset, has not been released yet as of the README. The project is supported by the US Army Research Laboratory and DSTA.