Michael S. McDaniel

Disparity space image's features analysis for error prediction of a stereo obstacle detector for heavy duty vehicles

Vision-based perception has been explored as low-cost, flexible technology for industrial applications and ADAS. Its inherent flexibility presents a challenge quantifying performance and often even quantifying increases or decreases in system performance as conditions change. Experience enables designers to employ various ”rules of thumb” while commercially viable products require quantitative performance. This paper explores the correlation between features and characteristics of the Disparity Space Image (DSI) and resulting performance for an object detection application. The specific application is an object detection system suitable for highly chaotic environments often found in earthmoving industry. Features and characteristics with strong correlations can be used to improve system design and predict system performances at run-time. High-quality stereo images are used to characterize baseline system performance. These images are then artificially degraded to simulate fog, darkness, and blurring and subsequent system performance compared to baseline results.