Semantic Segmentation of Driving Environment in Fog Conditions Using Pseudo-Fog Images

Image recognition using a camera can recognize the necessary information for driving, such as object recognition and tracking using deep learning technology, white line detection, signal recognition, sign recognition, and driving area identification. However, the actual camera image often deteriorates due to bad weather such as fog, haze, and dust. We propose pseudo-fog images data augmentation by depth estimation using semantic segmentation databases to improve recognition in fog environment. This method estimates the distance to cars in images from the label image of the databases and creates pseudo-fog images using the fog model formula. This method makes it possible to easily create fog databases from the original databases