Real World Semantic Occupancy Prediction for Advanced Air Mobility

Perception of the environment represents a foundational yet intricate challenge for autonomous robotic systems. Recent computer vision and deep learning advancements have established semantic occupancy grids as a standard geometric scene perception and semantic understanding framework. Numerous deep learning models have been trained and validated extensively on benchmark datasets, significantly augmenting autonomous navigation, path planning, and decision-making capabilities. However, most of these datasets are tailored specifically for terrestrial autonomous driving applications, with very limited consideration of aerial mobility scenarios.
A novel semantic occupancy dataset has been introduced in this research, specifically designed for aerial mobility, and developed using a novel data generation pipeline that relies solely on monocular RGB aerial imagery. The pipeline integrates 3D reconstruction, semantic fusion, and voxel densification to produce dense voxel-based representations of the environment enriched with visibility-aware semantic labels. Annotation time, which traditionally requires weeks per scene, is significantly reduced to mere hours through its autonomous operation with minimal manual intervention. [...]