Kentsel dünyanın 3D algısı için derin öğrenme tabanlı tespit ve segmentasyon

The rapid development of autonomous driving technologies has highlighted the growing need for perception systems that can understand and interpret complex urban traffic scenes. This thesis presents a deep learning-based, vision-only framework that focuses on the detection and segmentation of three key elements in traffic environments: roads, vehicles, and pedestrians. The system operates using 2D image projections generated from 360-degree cameras, without relying on LiDAR or any other external sensors. This approach aims to offer a scalable and cost-effective solution for intelligent transportation applications. To train and evaluate the proposed models, a custom dataset was created using Google Street View imagery collected from four major European cities: Istanbul, Paris, Munich, and Marseille. The dataset includes 8,932 labeled images and more than 149,000 object annotations, providing a diverse range of traffic scenes under varying urban conditions. YOLOv8 and YOLOv10 models were used for object detection, while DeepLabV3 was applied for semantic segmentation. The models were evaluated across different train-validation splits using standard metrics. YOLOv10 achieved the best detection performance with a mAP@0.5 score of 0.685 and an overall precision of 0.76. DeepLabV3 produced strong segmentation results, including IoU scores above 0.85 for road detection and F1 scores exceeding 0.80 in clearly defined object regions. These results indicate that image-only systems can effectively perform traffic scene analysis in real time. The framework developed in this thesis demonstrates the potential of deep learning for urban scene understanding and contributes a new dataset that supports future research in vision-based autonomous navigation.