Semantic segmentation enables pixel-wise classification of visual scenes, which is essential for autonomous vehicles to perceive their surroundings correctly. In order to address the challenges of intricate driving environments, we present an attention-augmented U-Net model in this work. Accurate segmentation makes it possible to precisely identify objects like roads, cars, pedestrians, and traffic signs for tasks like lane navigation, obstacle avoidance, and making safe decisions. Traditional segmentation techniques often suffer from occlusion, small object detection, and weather and lighting changes. By integrating attention mechanisms into the U-Net architecture, the U-Net Attention model was developed to get around these limitations. This enhanced the model’s ability to focus on relevant features and increased segmentation accuracy. The model was trained on the CamVid dataset and evaluated using the Adam optimizer over 50 epochs. The results outperformed more traditional models such as SegNet and ENet, with an overall accuracy of 90.42% and a mean Intersection-over-Union (mIoU) of 72%. Despite its promising performance, the model had trouble with small or highly obscured objects and changes in the environment, highlighting the need for further advancements such as hybrid architectures or self-attention modules. Future research will focus on improving the model for real-time deployment on edge devices and merging multiple datasets to improve segmentation performance in situations involving rare objects and extreme weather in order to further boost the robustness and dependability of autonomous driving systems.