A Patch Based Fisheye Image Segmentation for Satellite State Characterization

Corentin Menier, Cyril Meurie, Yassine Ruichek, Juliette Marais

Abstract: GNSS algorithms have shown wide flexibility in a number of domains, but still raises safety concerns in critical applications such as autonomous vehicles due to environmental impact on degraded satellite signals. In order to account for these degradations, previous methods rely on the accurate analysis of the receiver’s environment, with vision systems playing a major role in recent solutions.This study focuses on using vision for enhanced localization, relying on GNSS data for positioning via techniques like propagation time and triangulation. Over the past decade, significant research has addressed environmental corruption, particularly Non-Line-Of-Sight (NLOS) scenarios in urban canyons.Approaches integrating vision and GNSS sensors have gained popularity for their effective environmental characterization. Our method uses a sky-oriented wide-angle camera to segment the environment into sky and non-sky regions. However, the binary output can be restrictive, limiting the ability to handle uncertainty and instability in positioning.We propose a novel attention-based patch segmentation process for fisheye images, which allows the model to access both local and global information. An additional fuzzy class, introduced in previous work, is used to represent uncertain predictions and improve GNSS data processing.The localization pipeline is trained and tested on a publicly available dataset from the ISAE-SUPAERO laboratory, containing both images and GNSS data."