Photo-Realistic and Labeled Synthetic UAV Flight Data Generation Using ROS and Gazebo
Lucas Cruz da Silva, Juliana Quintiliano Ferreira, Paulo Alfredo Frota Rezeck, Michel M. Silva, Thiago Luis C. Gomes · 2024
Visually realistic simulations are pivotal in advancing state-of-the-art in robotics and automation by creating comprehensive image datasets for numerous applications. Synthetic datasets address challenges posed by the high costs of collecting, limited accessibility, and potential dangers associated with creating real image datasets. Even when real images are accessible, annotating them remains an expensive and labor-intensive process. This paper presents a novel framework for producing photo-realistic synthetic images with depth and semantic segmentation labels in a simulated UAV environment using ROS and Gazebo. By leveraging public orthophotos and classified lidar point clouds, our framework offers full control over simulation parameters, including sensor configurations, UAV models, environmental conditions, and autonomous flight. Additionally, the framework employs a coverage control sampling strategy to ensure representative and comprehensive synthetic data generation. Our experiments demonstrate the capability of the framework to generate highly realistic images, particularly in scenarios that are restricted or expensive for real-world data collection. The proposed approach significantly reduces the cost and labor associated with dataset creation, providing a valuable resource for the robotics and AI community.