SD Map Localization: A Deep Learning Approach
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Complex adaptive systems (MPCAS), MSc
Publicerad
2024
Författare
Abdul Rahuman, Sheik Meeran Rasheed
Jathavedan, Sameer
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Accurate localization is critical for the safe and efficient operation of autonomous
vehicles, enabling precise navigation and real-time decision-making. This thesis focuses
on improving Standard Definition (SD) map based localization, by leveraging
deep learning techniques. The research addresses key questions, including how to
optimally encode SD map data and sensor data, particularly the Global Navigation
Satellite System and Inertial Navigation System sensor, for deep learning, and train
models to perform accurate map matching and vehicle localization along the correct
road segment.
The thesis develops a deep learning-based localization framework for autonomous
vehicles, focusing on SD map data. It introduces three main components: a Polyline
Encoder using either Graph Neural Networks (GNN) or Transformers, a Map
Matching Network based on cross-attention, and a Point Prediction Network consisting
of a simple Multi-Layer Perceptron. The model encodes ego trajectories and
map links, matches the map data with the vehicle’s trajectory, and predicts precise
location. Our results show that the GNN consistently outperforms the Transformer
on both map matching and point prediction. The model’s performance varies based
on the training and testing data used, with the last point of the trajectory often
being sufficient for accurate localization. The study also compares the deep learning
model with classical algorithms and finds that the GNN-based localization model
significantly improves localization accuracy. Overall, our thesis demonstrates that
leveraging deep learning techniques, particularly GNN-based architecture for encoding,
along with cross-attention based architecture for map matching, has the
potential to significantly enhance SD map localization for autonomous vehicles.
Beskrivning
Ämne/nyckelord
Deep Learning, Map Localization, Transformer, Graph Neural Network