Writing · 2025
Gaussian Splatting for UAV Perception
A research proposal and technical survey on 3D Gaussian splatting as a UAV perception representation, with a benchmark design for safety fidelity and on-board cost.
Research proposal, Cal Poly Pomona · 2025
3D Gaussian splatting reconstructs a scene as a field of anisotropic Gaussians rather than a mesh or a neural field, which makes it fast to render and unusually well suited to a vehicle that has to reason about free space in real time.
The survey covers how splatting works, its history relative to photogrammetry and neural radiance fields, and a comparative analysis of where it wins and where it does not.
The proposed evaluation is deliberately safety-first rather than photometric. It measures minimum-distance error, clearance violation rate at 20 and 40 cm thresholds and corridor navigation success; geometry quality by Chamfer distance and F-score at 1, 2 and 5 cm; pose accuracy by ATE and RPE; and the operational cost — FPS distribution, P95 latency, VRAM use, power draw and thermal throttling events — with the aim of identifying stable real-time, balanced and quality operating configurations with documented performance envelopes.