Efficient Deep Learning for Mapping and Localization of Intelligent Vehicles

SAVE board

  • PhD position
  • Delft

Website TU Delft

The Intelligent Vehicles group at TU Delft is seeking a PhD candidate with an interest in performing cutting edge research in the area of self-driving vehicles, in collaboration with TomTom, a global leader in mapping and navigation products.
Currently, highly automated vehicles commonly rely on detailed 3D maps created with SLAM algorithms and LIDAR data for accurate self-localization. However, these representations do not scale, are sensitive to changes in the environment, are sensor specific, and also computationally intensive. TomTom’s research product RoadDNA takes an alternative approaches to represent the road environment more efficiently. However, it uses a hand-engineered representation, which is mainly target at highway environments sensed with LIDAR at a fixed compression rate.
This PhD student will instead develop optimized representations for mapping and localization in complex urban environments by learning robust semantic feature representations through end-to-end weakly-supervised deep-learning. The novel methods can additionally support rough priors provided by GPS, structural priors from aerial imagery and existing map data, or even temporal context. A learned representation can thus focus on features which matter most in the local area, and henceforth reduce its size, and increase localization efficiency. Higher-level features are additionally more robust against environmental changes, and could be transferred between multi-modal sensor setups or multiple viewpoints.

To apply for this job please visit intelligent-vehicles.org.