مريم اسعدساماني

This thesis explores controlling autonomous vehicles in adverse weather conditions, such as snow, rain, fog, and nighttime lighting. The proposed approach involves Generative Adversarial Networks (GANs) to reconstruct images captured by the vehicle's camera. GANs are deep learning models comprising two neural networks: a generator and a discriminator, which are trained through adversarial processes. This model operates independently of the type and severity of weather conditions, reconstructing input images in real-time. During training, the network utilizes image labels, but it operates without them during eva‎luation. In other words, the network's deployment is independent of user-defined labels. Furthermore, the network's performance is maintained when encountering a combination of various adverse weather conditions. To improve accuracy and minimize the network's dependency on object orientations, Capsule Networks (CapsNet) have been integrated into the generator's architecture. Additionally, the loss function has been refined to maintain the details of the original image. A comprehensive dataset of images under various weather conditions has been prepared as the training data. The proposed method has been eva‎luated on BDD100k dataset, tested using the AVIS ENGINE simulator, and the results highlight the designed network's high accuracy and speed. This development represents a significant step forward in enhancing autonomous systems, facilitating safer deployment of this technology across diverse environmental conditions.

خودروي خودران , شبكه‌هاي مولد متخاصم , بينايي ماشين , كنترل هوشمند

Autonomous Vehicle , Generative Adversarial Network , Computer Vision , Intelligent Control

Maryam Asad Samani

Dr. Mohammad Farrokhi