Road Lane Marker Localization from Airborne LiDAR Surveys for Deep Learning

Abstract

Road lane markers are key to many applications—among, autonomous driving and infrastructure monitoring. Remote sensing data such as multispectral satellite or airborne imagery serve as basis to efficiently generate georeferenced maps on large scale. Supervised deep learning models for localizing of lane markers in overhead imagery requires manual annotation of many training data—a human labor-intensive and time consuming process. Weakly-supervised methodologies provide a potential solution to reduce or even eliminate the effort required in creating manual labels. Recently, the generation of noisy labels from high quality airborne laser surveys rendered convenient to train deep neural networks for identification of buildings, roads, and vegetation. The concept of AutoGeoLabel employs simple, physics-based rules to generate noisy, georeferenced semantic segmentation maps. In this thesis we expand and study the AutoGeoLabel approach for lane markers given their statistics of LiDAR reflection information. Curating a novel lane marker benchmark data set for a total of about 15 km 2 of major roads in New York City from Google Maps co-registered with noisy AutoGeoLabel-annotated LiDAR-statistics lane marker signals and about 4% of vectorized ground truth labeling, our initial set of experiments indicate that the training of a vanilla U-Net seems challenging given the subtle signal of lane markers in RGB overhead imagery.