Toward Appearance-based Autonomous Landing Site Identification for Multirotor Drones in Unstructured Environments

TL;DR

The paper tackles autonomous landing site identification in unstructured environments using appearance-based RGB image segmentation to avoid heavy sensing and GPS-based constraints. It introduces a pipeline that reconstructs 3D terrain from terrain surveys, automatically generates synthetic, labeled RGB datasets, and trains a small U-Net tailored for onboard inference on a Google Coral TPU. Key contributions include the data-generation workflow, a compact real-world validation dataset, and demonstration of real-time onboard performance with up to 0.83 accuracy on 10-second validation videos. The approach offers a practical path toward robust, low-overhead autonomous landing for multirotor drones in diverse environments, while acknowledging sim-to-real gaps and the need for broader data and architecture exploration.

Abstract

A remaining challenge in multirotor drone flight is the autonomous identification of viable landing sites in unstructured environments. One approach to solve this problem is to create lightweight, appearance-based terrain classifiers that can segment a drone's RGB images into safe and unsafe regions. However, such classifiers require data sets of images and masks that can be prohibitively expensive to create. We propose a pipeline to automatically generate synthetic data sets to train these classifiers, leveraging modern drones' ability to survey terrain automatically and the ability to automatically calculate landing safety masks from terrain models derived from such surveys. We then train a U-Net on the synthetic data set, test it on real-world data for validation, and demonstrate it on our drone platform in real-time.

Figures (5)

• Figure 1: One of three data collection sites. We manually and randomly picked 18 validation sites, and we number them as follows, marking safe with an S and unsafe with a U: (1U) -- an archery target on a soccer field, (2U) -- a bush, (3S) -- a flat, dirt area, (4U) -- a large, cracked rock mound, (5U) -- high vegetation area, (6S) -- flat, mossy area in a lava field, (7S) -- flat, grassy area, (8S) -- flat, mossy area in a lava field, (9U) -- crack in a lava field, (10S) -- dirt patch in a lava field, (11S) -- road, (12U) -- person, (13U) -- very rough lava field, (14S) -- model aircraft runway, (15U) -- sloped, gravel edge of a soccer field, (16S) -- green spot in a soccer field, (17S) -- middle of a soccer field, (18U) -- soccer goal. Map source: Loftmyndir ehf. loftmyndir
• Figure 2: Example of manual segmentation of a river in the summer house dataset by isolating the river and adding a manual classification to it in CloudCompare. It is not feasible to isolate the river by simply filtering on the altitude above sea level (ASL) since the terrain has a significant slope.
• Figure 3: Pipeline for creating labeled image datasets from terrain surveys.
• Figure 4: Example images and masks from the synthetic data set.
• Figure 5: Example predictions with and without post-processing