Recognition and Reconstruction of Road Marking with Generative Adversarial Networks (GANs)

Abstract

Road markings play an essential role in road safety and are one of the most important elements to guide autonomous vehicles and help the driver on the road. The recognition and detection of road markings has been very successful in recent years with the rapid development of deep learning technology. Although considerable work and progress has been made in this area, they often depend excessively on unrepresentative datasets and inappropriate conditions. In this article, to overcome these drawbacks, we propose a deep learning system for the extraction, classification, and completion of road markings which generates high-quality samples for data augmentation. For this, an in-depth learning network proposed to successfully recover a clean road marking of a fuzzy route using generative contradictory networks (GAN). The proposed data augmentation method, based on mutual information, can preserve and learn the semantic context from the actual dataset. We build and train a GAN model to increase the size of the training dataset, which makes it suitable for recognizing the target. Our system can generate clean samples from fuzzy samples and surpasses other methods, even with unconstrained road marking datasets.