Abstract
Simulation and synthetic data generation for AI training in aerospace domain come in various forms. To achieve both tasks effectively, constructing a virtual environment with high compatibility is essential. The environment must cover a vast area with individual objects that allow vehicles to navigate and interact with their surroundings. High quality and reality are necessary to capture every scene as synthetic data. Furthermore, it must be thoroughly optimized for efficient and seamless operation.
Among all components of the environment, buildings pose a significant challenge in meeting these requirements. Buildings appear in diverse shapes, materials, and architectural styles depending on the region. As essential structures in human environments, they occupy a large portion of the scene. Their complexity can lead to performance issues, such as reduced frame rates or system instability. Despite the importance, there is no widely adopted method to resolve these problems.
This paper proposes a method for the automatic generation of building 3D meshes in LOD3 detail over large regions. By integrating BIM and GIS concepts, the proposed approach produces realistic, optimized 3D building models with a minimal amount of information. These meshes require less computing power compared to other 3D building models of similar quality while remaining as individual objects for detailed simulation purposes. The study demonstrates the effectiveness of this approach through qualitative and quantitative analysis. Scenes of the virtual environment corresponding to the actual region are presented as an example. Additionally, evaluation metrics of AI model trained with the generated synthetic data from the virtual environment shows the method’s effectiveness.