Map Matching (2025)

MAP MATCHING using deep neural networks in a multi-modal environnement

The subject of my internship is the development of a navigation module capable of accurately locating a drone equipped with an on-board camera. The latter is also equipped with a GPS sensor giving its absolute position on Earth. However, signal loss can occur during flight. To maintain an estimate of the drone’s position in the event of GPS loss, conventional mobile robotics methods exist, notably visual odometry algorithms, but these suffer from time drift due to relative rather than absolute estimates. This report looks at an absolute localization solution : map registration. The aim is to use a satellite map of the area overflown, and the image acquired at a given moment, to estimate the position of the drone. A major constraint is the need to fly at low altitude in a rural context.

First, I reviewed the state of the art in aerial mapping, to understand the different families of approaches, as well as the limitations of each. I evaluated their relevance to our use case, and for the open-source implementations,

I tested their execution on in-house data. Based on these results, I worked on a proof-of-concept that would meet our needs and take into account the project’s constraints.

Due to confidential reasons, no other explainations are provided.