Camera image-quality assessment for autonomous driving perception under lens variation and winter weather conditions
Аннотация
This study evaluates camera-input image quality relevant to autonomous-driving perception using controlled chart-based imaging, system-level camera–lens configuration comparison, quantitative winter-weather measurements, and Imatest-based image-quality analysis. Two complementary experiments were conducted. First, two camera–lens imaging configurations were evaluated using the same Allied Vision Mako G-319C camera. Configuration 1 used a 25 mm lens at a 5 m camera-to-target distance, whereas Configuration 2 used a 16 mm lens at a 3 m camera-to-target distance. Under the tested acquisition settings, Configuration 2 showed approximately 29.4%, 35.3%, and 36.8% higher mean weighted MTF50, MTF30, and MTF20 values, respectively, and approximately 37.5% lower mean weighted chromatic aberration than Configuration 1, while Configuration 1 showed slightly better color accuracy. Because focal length, camera-to-target distance, and exposure time differed between the two configurations, these results represent system-level imaging-performance differences rather than isolated lens-optics effects. Second, camera-input image quality was evaluated outdoors under light snow, snow, heavy snow, and heavy snow/small hail using a fixed camera–lens configuration. ASDO/Parsivel 2 measurements were used to quantitatively characterize the outdoor conditions, and four chart images per weather condition were analyzed to assess frame-to-frame consistency. The repeated-image analysis showed a non-monotonic image-quality response across the evaluated winter-weather conditions. Snow produced the highest mean MTF50 and MTF20 values, whereas heavy snow/small hail produced the lowest mean MTF50 and MTF20 values. Chromatic aberration followed a different pattern, with heavy snow producing the highest mean CA area and the greatest frame-to-frame variability in this metric. Overall, the results show that camera-input image quality varies across imaging configurations and measured winter-weather conditions, supporting the combined use of quantitative weather measurements and repeated image-quality assessment for characterizing camera inputs relevant to autonomous-driving perception.