Lane Centerline Detection from Orthophotos using Transformer Networks
Name
Karl-Johan Pilve
Abstract
A high-definition (HD) map containing accurate lane network data is essential for autonomous driving. Yet, the creation of an HD map for a larger area is often a time-consuming and labor-intensive process. Recently, neural networks utilizing the transformer architecture have shown promising results in the field of computer vision. One such example is RNGDet, which uses aerial images to iteratively generate a road network graph. This thesis explores the viability of fine-tuning the RNGDet model with lane network data to generate a lane network graph for the entire city of Tartu based on high-resolution orthophotos. The obtained results show that RNGDet could be used in principle to generate a lane network graph. However, the model's architecture likely requires more extensive modifications to accommodate the differences between road and lane network data. Since the orthophotos may not contain all the necessary information for correct lane network generation, the best-performing model used orthophotos with rasterized raw GPS trajectories, which were collected by a survey vehicle. The obtained results also indicate that additional training data is needed to increase the quality of the generated lanes enough that they could be used in HD maps.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Tambet Matiisen, Edgar Sepp
Defence year
2024