What did you find out?

We developed an Image Quality Safety Model (IQSM) that accurately detects and diagnoses image quality issues that impair object detection performance, achieving approximately 98% accuracy in predefined hazardous scenarios. The scope of the paper is highly automated agricultural machines (HAAM), such as autonomous tractors. These vehicles, operating in open fields, may encounter various objects—such as unexpected human beings—that need to be correctly identified. A sensor system failure in this context can lead to hazardous situations. Our specific focus is on camera systems, which may experience environmental disturbances on the lens, such as dirt, raindrops, or fog. Just as a person wearing dirty glasses may struggle to see clearly, the perception system's performance is impaired—posing a safety risk.

What challenges did you face during the project?

We faced challenges in generating realistic low-quality images to represent hazardous scenarios and in modeling the complex relationship between image quality and AI model performance.

Why are these results important for practice?

These results enable safer and more reliable perception systems in highly automated agricultural machines by detecting unsafe input conditions in real time and supporting timely mitigation actions.

Publication

Lee, Changjoo; Schätzle, Simon; Lang, Stefan Andreas; Maier, Michael; Oksanen, Timo; Fault Management System for the Safety of Perception Systems in Highly Automated Agricultural Machines; ICRA 2025