Implicit representations for place recognition and robot localisation

Overview

This project will develop a novel localization pipeline based on implicit map representations. Unlike traditional approaches that use explicit representations like point clouds or voxel grids, the map in our project is represented implicitly in the weights of neural networks such as Neural Radiance Fields (NeRF). You will get a chance to develop a new class of localization algorithms that work directly on the implicit representation, bypassing the costly rendering step from implicit to explicit representation. The designed algorithms will address the still difficult problem of accurate large-scale localization under challenging environmental conditions, and in situations where GPS is not reliably available, which is crucial for autonomous and assisted driving.

Contact

Contact the supervisor for more information.