An international research team has released a publicly available image dataset designed to enhance visual navigation for autonomous photovoltaic cleaning robots. The collection includes 6,200 annotated photos of solar panels, featuring pristine images, synthetically added contamination at five severity levels, and laboratory captures of physically soiled modules. By maintaining consistent reference-line geometry across varying soiling conditions, the dataset enables systematic testing of robotic vision algorithms for reliable navigation cue detection.
Images were captured with an HF868-2 camera mounted 15 cm above the panel surface on a cleaning robot, using a 2.7 mm wide-angle lens with a 135° field of view. Recorded at 320 × 240 pixels and 60 fps, the footage was geometrically calibrated and downscaled to 160 × 120 pixels to suit the limited processing power of onboard systems. Researchers manually annotated 800 clean images, providing a foundation for future AI-driven navigation solutions.
Source: pv magazine Global
