Attention-Based Deep Recurrent Neural Network for Semantic Segmentation of 4-D Radar Data Acquired During Landing Maneuver

Abstract

Autonomous driving vehicles are being more and more popular in the community with the rise of artificial intelligence systems. However, in the context of airborne navigation, it remains a challenge, especially during landing maneuver. In order to operate in all conditions (weather, day, and night) and in all airports, we propose a runway localization method based on images acquired by an onboard radar. The proposed algorithm is a radar data segmentation method designed for use by an aircraft, as an on-board system, to provide the pilot, whether human or automatic, with a runway location prediction to facilitate and secure the landing maneuver. This article describes the acquisition and labeling of a large-scale real dataset over 18 airports in France and Switzerland, and the proposition of an attention-based deep recurrent neural network (RNN) for semantic segmentation of 4-D radar data acquired during a landing maneuver. This end-to-end trainable neural network combines attention mechanisms adapted to the geometry of an approach scene, with the exploitation of spatial-temporal information via recursive cells, all being associated with a convolutional segmentation model (patent pending). This article proposes a sensitivity analysis of Lyon’s airport to tune the hyperparameters, demonstrating the interest in adapting the attention sequence, especially through the shape of patches. The experimental results have shown the benefit of each block in the model. Extensive experiments on the other available airports have allowed validating the potential of the proposed network. Experiments have shown a considerable gain of about 0.17 on the DICE score associated with the exploitation of attention mechanisms and recursive cells and a gain of 0.1 compared to the SegFormer-B0 model.