Lane-aware deep learning for road lane detection in rain (part 2)

Abstract

We presented a solution for lane recognition in rainy weather conditions in our study that incorporates a self-attention module into an ENet model architecture. We concentrate on specific weather conditions, such as rain, because they have received less attention as a result of the numerous obstacles they present. One of the primary reasons is that rain water distorts road lines, resulting in the appearance of distorted lanes and markings. We trained our model on annotated labelled data obtained from the author of VPGNet, which comprises a variety of situations, including ones in rainy conditions. We presented a deep learning technique in our ENet architecture, utilizing ENet semantic segmentation and a self-attention module. Our model can recognize and classify a given image on a pixel-by-pixel basis, with each pixel representing a distinct class. We developed a robust model that is capable of performing effectively under a variety of real-world scenarios while maintaining a frame rate of 20 frames per second. Our testing findings indicate that our technique reaches a level of accuracy comparable to that of the fundamental ENet architectural model.