هدي رضانيا

This thesis examines the challenges of semantic segmentation of foggy images in autonomous driving systems. Autonomous vehicles require high accuracy in understan‎ding their surroundings under various weather conditions. However, fog, by reducing visibility, poses one of the biggest obstacles to the proper functioning of these systems, making the detection an‎d precise segmentation of objects an‎d their boundaries in foggy images significantly more difficult than in clear weather conditions. This issue threatens the safety of autonomous vehicles an‎d hampers their efficiency in specific situations. In this research, advanced deep learning methods are employed to enhance semantic segmentation of images under foggy conditions. Considering the high cost of collecting foggy datasets for autonomous vehicles, domain adaptation methods are utilized. One of the key approaches in foggy image semantic segmentation is unsupervised domain adaptation (UDA), which allows models to adapt to foggy images without labeled target data. This approach improves the accuracy of deep learning models by reducing the gap between the source an‎d target domains. Additionally, the ACDC dataset, one of the most renowned datasets for foggy images in autonomous vehicles, containing urban imagery, is utilized. Furthermore, the Masked Image Consistency (MIC) model is introduced, leveraging spatial contextual relationships to enhance segmentation accuracy. By creating consistency between masked image predictions an‎d pseudo-labels, this model significantly improves the ability of models to accurately predict masked regions. Specifically, in domain adaptation from the Cityscapes dataset to ACDC for semantic segmentation, the MIC model achieves a 2.4% improvement in the mIoU accuracy metric compared to state-of-the-art models like HRDA. This advancement plays a crucial role in enhancing the safety an‎d efficiency of autonomous vehicles in foggy weather conditions.

قطعه‌بندي معنايي تصوير, تصاوير مه‌الود,خودروخودران,تطيق دامنه نظارت نشده

semantic segmentation,autonomous vehicles,foggy images,unsupervised domain adaption

hoda rezania

hadi shahriyar shahhosseini