Improving Autonomous Vehicles Maneuverability and Collision Avoidance in Adverse Weather Conditions Using Generative Adversarial Networks

Abstract

In recent years, there has been a significant increase in the development of autonomous vehicles. One critical task for ensuring their safety and dependability, is obstacle avoidance in challenging weather conditions. However, no studies have explored the use of data augmentation to generate training data for Deep learning (DL) models aimed at navigating obstacles in extreme weather conditions. This study makes a substantial contribution to the field of autonomous vehicle obstacle avoidance by introducing an innovative approach that utilizes a Generative Adversarial Network (GAN) model for data augmentation, with the objective of enhancing the accuracy of DL models. The use of a GAN model to generate a training dataset and integrate images depicting challenging weather conditions has been pivotal in enhancing the accuracy of the DL models. The extensive training dataset, consisting of 64,336 images, was created using three cameras installed in VSim-AV, an autonomous vehicle simulator, thereby ensuring a diverse and comprehensive dataset for training purposes. Three DL models (ResNet50, ResNet101, and VGG16 transfer learning) were trained on this dataset both before and after applying the data augmentation techniques. The performance of the augmented models was evaluated in a real-time environment using the autonomous mode of the VSim-AV simulator. The testing phase resulted in the highest accuracy rate of 97.2% when employing Resnet101 following the implementation of GAN. It was observed that the autonomous car could navigate without any collisions, showcasing a remarkable reaction time of 0.105 seconds, thus affirming the effectiveness of the approach. The comparison between the original and augmented datasets demonstrate the originality and value of this study, showcasing its significant contribution to the advancement of autonomous vehicle obstacle avoidance technology. This paper makes significant advances to the field of autonomous vehicle navigation by exploiting Generative Adversarial Networks (GANs) to improve obstacle avoidance capabilities in severe weather conditions, hence increasing safety and dependability in real-world applications.

Keywords

• Autonomous-vehicles
• obstacle-avoidance
• avoiding collision
• VSim-AV
• deep learning (DL)
• generative adversarial network (GAN)