Approches de Deep Learning pour la segmentation d’image en temps réel dans la conduite autonome

Abstract

Real-time semantic segmentation is a key task in computer vision, especially for autonomous driving systems where quick and accurate perception is essential. This thesis explores deep learning techniques for this task using the U-Net architecture, applied to the Cityscapes dataset of urban driving scenes. To enhance performance, we replaced the original U-Net encoder with EfficientNet-B0, aiming to achieve both high segmentation accuracy and real-time inference speed. The proposed model reached over 61.8% mean IoU and 128 FPS, making it highly suitable for real-time applications. The work was developed using Google Colab and benefited from an internship at CRTI, which provided technical support and valuable resources. Despite certain challenges such as limited annotation in the test set, the final results demonstrate the model’s efficiency and robustness.