<italic>ARD&#x00C1;N</italic>: Automated Reference-Free Defocus Characterization for Automotive Near-Field Cameras

Abstract

Measuring optical quality in camera lenses is crucial in evaluating cameras, especially for safety-critical visual perception tasks in automotive driving. While ground-truth labels and annotations are provided in publicly available automotive datasets for computer vision tasks, there is a lack of information on the image quality of camera lenses used for data collection. To compensate for this, we propose an Automated Reference-free Defocus characterization for Automotive Near-field cameras (ARDÁN) to evaluate Slanted Edges for ISO12233 in five publicly available automotive datasets using a valid and invalid region of interest (ROI) selection system in natural scenes. We use the mean of 50% of the Modulation Transfer Function (MTF50) in three Camera Radii (CaRa) segments and the Overall Spread in Heatmaps (O’SHea) for an $8\times 5$ distribution to evaluate the quality of edges in natural scenes. From the experiments performed, lenses with uniform spatial domains (i.e. little distortion) showed that MTF50 was constant between (0.18-0.25cy/px). With image rectification on the same scenes, MTF50 results artificially increased, no longer representing the camera lens. In contrast, for strong radial distortion, MTF50 varied extensively across the spatial domain between (0.12-0.4cy/px), where, in particular, Woodscape gave the highest average of MTF50 per region for natural scenes.

Keywords

• Automotive
• camera radii (CaRa)
• energy inspection limit above Nyquist (EiLíN)
• natural scenes
• optical quality
• the overall spread in heatmaps (O’SHea)