Vision Transformer-Based Autonomous Safe Landing Zone Detection for MAVs
TL;DR
This work systematically evaluates the proposed approach on aerial RGB images from a public dataset as well as on noisy depth images captured onboard a drone using a lightweight depth camera, showing that the ViT-based model consistently outperforms widely used CNN architectures, including ResNet, particularly in low-data regimes where generalization is crucial.
Abstract
We propose a novel methodology for autonomous landing zone detection in Micro Aerial Vehicles (MAVs) based on Vision Transformers (ViTs). The core contribution of this work lies in demonstrating that transformer-based architectures, originally developed for large-scale vision tasks, can be effectively adapted to safety-critical aerial robotics applications with limited training data. Unlike traditional Convolutional Neural Networks (CNNs), ViTs leverage self-attention mechanisms to model long-range spatial dependencies, enabling a more holistic understanding of scene geometry and surface suitability for landing. We systematically evaluate the proposed approach on aerial RGB images from a public dataset as well as on noisy depth images captured onboard a drone using a lightweight depth camera. Our results show that the ViT-based model consistently outperforms widely used CNN architectures, including ResNet, particularly in low-data regimes where generalization is crucial. Notably, the transformer model maintains strong robustness even when operating on degraded depth inputs. In addition to accuracy improvements, the proposed system achieves real-time performance, with an average inference time of [Formula: see text] ms on legacy GPU hardware. These findings highlight the practical feasibility and effectiveness of Vision Transformers for reliable, efficient MAV landing zone detection.