SkyShield: Occupancy as a Safety Interface for Low-Altitude UAV Autonomy

Researchers introduce SkyShield, the first monocular semantic occupancy benchmark for low-altitude UAV autonomy below 20 meters, addressing a critical gap in aerial safety perception. The dataset includes 36K annotated samples with 6-DoF pose tracking and a new safety-aware evaluation metric (KAR-mIoU) that prioritizes collision-critical risks over traditional accuracy measures.

SkyShield addresses a fundamental safety challenge in autonomous drone operations that has been largely overlooked by existing research. Low-altitude urban flight presents unique perceptual demands—vegetation, occlusions, and thin geometry determine navigability in ways that ground-level datasets and high-altitude aerial benchmarks cannot capture. Current UAV datasets rely on 2D annotations or 3D bounding boxes, inadequate for the continuous spatial understanding required for safe autonomous flight in constrained airspace.

The research community has primarily focused on driving datasets and high-altitude remote sensing, creating a perception gap where the most operationally challenging flight regime lacks proper benchmark infrastructure. SkyShield bridges this by providing frame-wise 6-DoF pose annotations and dynamic camera geometry—crucial for understanding how sensor perspective changes relative to static and moving obstacles. This reflects broader trends in embodied AI where static annotations prove insufficient for dynamic agents with changing viewpoints.

The introduction of KAR-mIoU represents significant methodological progress. Rather than treating all prediction errors equally, this metric weights voxels by kinematic reachability and time-to-collision, directly aligning evaluation with safety outcomes. This transforms occupancy prediction from an abstract perception task into a concrete safety problem, making it directly relevant to autonomous system certification and deployment decisions.

The practical implications extend beyond academic benchmarking. As commercial UAV operations expand into congested urban environments, safety-critical perception systems require evaluation frameworks that reflect real operational constraints. SkyOcc's geometry-first baseline and safety-prior optimization demonstrate how domain-specific knowledge improves both safety and efficiency. Future development will likely focus on scaling these methods to real-world sensors and integrating with flight controllers.

• →SkyShield is the first monocular occupancy benchmark specifically designed for low-altitude UAV flight below 20 meters in urban environments
• →KAR-mIoU metric reweights prediction errors by kinematic reachability and collision timing, directly measuring safety-critical performance rather than general accuracy
• →The dataset provides 36K samples with frame-wise 6-DoF pose and dynamic camera geometry, capturing the unique perceptual demands of aerial autonomy
• →Existing driving and high-altitude datasets miss the defining challenges of human-scale airspace navigation with thin geometry and dense occlusions
• →SkyOcc baseline demonstrates that geometry-first approaches with safety priors outperform conventional occupancy networks for autonomous flight applications