Semantic Segmentation for Road Surface Detection in Snowy Environment
Sirawich Vachmanus, Ankit A. Ravankar, Takanori Emaru, Yukinori Kobayashi
Abstract
Inclement weather conditions such as snow can severely hamper visibility and make it more challenging to recognize the obstacles on the road. This situation is very challenging for drivers to locate the drivable regions on the road. For autonomous driving vehicles, inclement weather conditions such as snow are a severe challenge due to poor visibility, lack of features on the ground, and slippery surfaces on roads. This research attempts to classify the road and obstacle regions on the snowy roads for the application of autonomous driving. The dataset using in this research are based on RGB images from such conditions. It includes different image resolutions that were collected by multiple camera systems such as USB camera, dashcam, and a smartphone camera. A deep learning network employing the structures are proposed in this research. The proposed structure uses the atrous convolution network and pyramid supervision for extracting features from snow road images. The proposed method can reach 86.0% mean ratio of the intersection with a little difference in the processing time. The boundary characterization of the proposed method has higher efficiency than other deep learning networks and also gained the highest classification performance when compared with existing techniques.