Title: SkyShield: Establishing Occupancy as the Safety Layer for Low-Altitude UAV Autonomy

Abstract:

In the realm of low-altitude Unmanned Aerial Vehicle (UAV) autonomy, 3D spatial comprehension serves not just as a perception goal, but as the critical safety bridge connecting human directives with physical flight operations. For urban airspace situated below 20 meters, the navigability of an aerial agent hinges on its ability to interpret thin geometries, occlusions, vegetation, and urban clutter. However, current UAV datasets predominantly rely on 2D annotations or 3D bounding boxes. Furthermore, driving-oriented occupancy benchmarks are designed around stable, ground-level sensor configurations. Both approaches fail to capture the unique dynamics of low-altitude flight: a moving aerial body utilizing a front-facing monocular camera to observe free and occupied spaces, characterized by frame-by-frame variations in 6-DoF pose and camera extrinsics.

To address this deficiency, we present SkyShield, which we believe is the inaugural front-view monocular semantic occupancy benchmark tailored for urban UAV operations under 20 meters. Leveraging the CARLA simulator, SkyShield comprises 36,000 front-view UAV samples spanning various urban environments and weather scenarios. Each image is synchronized with frame-specific 6-DoF UAV poses, dynamic camera geometry, UAV states, and front-frustum semantic occupancy labels. Additionally, we introduce KAR-mIoU, a UAV-specific and dynamics-sensitive metric that adjusts voxel-level assessments based on kinematic reachability and time-to-collision. This approach exposes safety-critical hazards that traditional mIoU metrics often overlook. To address the complexities of this new domain, we offer SkyOcc, a geometry-centric monocular baseline. SkyOcc incorporates frame-wise UAV attitude into its projection process, merges temporal occupancy features, and employs safety-prior optimization to maintain sparse, collision-critical structures. Collectively, SkyShield, KAR-mIoU, and SkyOcc redefine occupancy as the foundational safety interface for low-altitude aerial autonomy. The code and dataset will be made publicly available.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC