UF researchers warn that simple repeating visual patterns can disrupt depth perception in autonomous systems that rely on stereo cameras, potentially causing unsafe maneuvers or collisions.

The work emphasizes addressing foundational sensing and depth-estimation weaknesses to improve resilience against both intentional attacks and unpredictable real-world conditions.

The vulnerability affects a range of technologies including self-driving cars, drones, and ground robots, indicating it is not limited to a single manufacturer or application.

The study tested multiple sensors, algorithms, and AI models across collaborators, finding a shared underlying weakness in depth estimation when encountering repeated patterns.

Findings will be presented at the ACM Conference on Computer and Communications Safety (ACM CCS) in November, with practical testing conducted in UF’s autonomous-vehicle facility.

The issue can arise from both deliberate attacks and natural scene patterns, underscoring a broader safety risk in AI-driven perception systems.

Defenses were developed that target the root cause rather than merely increasing training data: traditional depth-estimation algorithms gained a software-based change to recognize repeated patterns and AI models were adjusted to respond differently to such patterns.

Experiments showed that patterns like checkerboards or striped textures can make parts of an object appear closer or farther, and a small pattern can be sufficient to influence depth estimates.

Researchers demonstrated a real-world driving scenario where a back-side pattern caused an autonomous vehicle to brake, illustrating potential hazards for following traffic.