Can we prove vision AI is safe against camera motion?

Deploying vision neural networks in safety-critical systems like autonomous vehicles requires mathematical proof they won't fail under real-world perturbations. Current verification only handles simple pixel-level noise, not the geometric distortions caused by camera motion. This work derives tight mathematical bounds on how camera movement warps images of planar scenes (road markings, runways, workspaces) by analyzing the continuity of perspective transforms. They achieve 89% speedup over prior methods and expose vulnerabilities in a real runway-classifier system. Code released.