Fast shadow detection for urban autonomous driving applications

This paper presents shadow detection methods for vision-based autonomous driving in an urban environment. Shadows misclassified as objects create problems in autonomous driving applications. Real-time efficient algorithms in dynamic background settings are proposed. Without the static background assumption, which was often used in previous work to develop fast algorithms, our scheme estimates the varying background efficiently. A combination of various features classifies each pixel into one of the following categories: road, shadow, dark object, or other objects. In addition to pixel level classification, spatial context is also used to identify the shadows. Our results show that our methods perform well for autonomous driving applications and are fast enough to work in real time.