U-Net Based Enhanced Lane Detection Learning With Directional Lane ROIs for Harsh Environments

Abstract

Recent advancements in artificial intelligence technology have propelled extensive research in the field of autonomous driving vehicles. Artificial intelligence's application in lane detection has effectively addressed challenges that were previously difficult to overcome with conventional techniques. This paper reduced the number of U-Net parameters required for learning to achieve faster processing. Additionally, it generates directional Edge images and incorporates them into the training to prioritize lane detection during ongoing driving. To ensure stable detection even in adverse conditions such as low-light situations, it employs a Bilateral Filter to suppress noise and increases the image's contrast using MSR (Multi Scale Retinex). The proposed method demonstrates greater stability, faster learning, and superior results compared to simple U-Net or 3-channel approaches.