Localization in Aerial Imagery with Grid Maps using LocGAN

Abstract

In this work, we present LocGAN, our localization approach based on a geo-referenced aerial imagery and LiDAR grid maps. Currently, most self-localization approaches relate the current sensor observations to a map generated from previously acquired data. Unfortunately, this data is not always available and the generated maps are usually sensor setup specific. Global Navigation Satellite Systems (GNSS) can overcome this problem. However, they are not always reliable especially in urban areas due to multi-path and shadowing effects. Since aerial imagery is usually available, we can use it as prior information. To match aerial images with grid maps, we use conditional Generative Adversarial Networks (cGANs) which transform aerial images to the grid map domain. The transformation between the predicted and measured grid map is estimated using a localization network (LocNet). Given the geo-referenced aerial image transformation the vehicle pose can be estimated. Evaluations performed on the data recorded in region Karlsruhe, Germany show that our LocGAN approach provides reliable global localization results.

Robust Self-Tuning Data Association for Geo-Referencing Using Lane Markings

“…Other approaches are based on learning methods [15], [16], [17], where a place is recognized from both aerial imagery and on-board sensors. Such strategies usually suffer from feature sparsity in the rural areas and depend strongly on rich-information environments, such as urban.…”