Impact of Windshield Optical Aberrations on Visual Range Camera Based Classification Tasks Performed by CNNs

Cameras operating in the visual range of electromagnetic spectrum are central to advanced driver assistance systems (ADAS). Front cameras, analyzing traffic, often located behind windshield detect and classify objects.Thus, area within camera’s field view is a part optical system. Simple windshields consist two curved glass surfaces connected by thermoplastic interlayer. Due defects present raw glass, as well those introduced during bending lamination process, will have aberrations. While quality may be suitable for human vision, it can fall short what needed machine vision. In this article we investigate how aberrations generated laminated safety (LSG) influence performance camera system based on this, classification image content convolutional neural network (CNN) affected. A method wavefront measurements LSG samples presented, which allows us parameterize linear model Zernike Space. From derive space-variant point spread functions (PSFs) apply dataset simulate windshield’s impact image. As use case, CNN was trained unmodified compared modified versions with models applied. We measured modelled different samples, one high other low quality. compare prediction accuracy data. The highquality sample had negligible effect overall accuracy, while low-quality lowered up ten percentage points due