Gaussian Splatting for UAV Perception

Gaussian splatting is an explicit, differentiable scene representation: the scene is modeled as N anisotropic 3D Gaussians, each with a center, a symmetric positive-definite covariance, an opacity, and a view-conditioned color expressed in spherical harmonics, rendered through a fast rasterizer. Because the primitives are explicit, they can be added, split and pruned during optimization — which is what gives the method its speed and its suitability for online updates on a vehicle.

The survey places it against what came before. Classical structure-from-motion and multi-view stereo, standardized by COLMAP, recover accurate metric geometry and meshes, but are texture-dependent and computationally expensive. Dense volumetric mapping — KinectFusion, then Voxblox and NVBlox extending TSDF to ESDF for planning — gives direct distance queries, which is exactly what a planner wants, at the cost of memory and compute that scale cubically and geometry that over-smooths.

Neural radiance fields reframed view synthesis as differentiable volumetric rendering through an MLP, and efficiency work followed quickly: hash-grid encodings in Instant-NGP, explicit grids in Plenoxels, tensor factorizations in TensoRF. They produce photorealistic, view-dependent appearance but train and render slowly, and extracting usable geometry from them is awkward. 3D Gaussian splatting unified explicit anisotropic primitives with differentiable rasterization to get photoreal rendering and fast optimization together, with geometry-focused variants such as GSDF, 2DGS and SuGaR attaching signed-distance or surface-aligned heads. Its weaknesses are pose sensitivity, variable geometry fidelity, and — the gap this proposal targets — almost no work on embedded performance.

The proposed evaluation is deliberately safety-first rather than photometric, because a UAV does not care how pretty the reconstruction is. It measures safety fidelity through minimum-distance error, clearance violation rate at 20 and 40 cm thresholds, and corridor navigation success; geometry quality through Chamfer distance and F-score at 1, 2 and 5 cm; and pose accuracy through ATE and RPE.

Alongside those it measures what the representation costs to run: FPS distribution, P95 latency, VRAM usage, power consumption and thermal throttling events. The output is meant to be a set of stable operating configurations — real-time, balanced and quality — each with a documented performance envelope and a recommended resource cap, so that a vehicle integrator can pick one knowing where it falls over.