Accidents resulting from road damage are becoming a serious concern, emphasizing the need for efficient monitoring systems and timely government intervention. This research highlights the potential of advanced AI-driven solutions in road safety management, providing a practical approach to efficiently monitoring and maintaining road conditions. It presents a real-time embedded vision system for automatic road damage detection using deep learning techniques. The system is designed to classify six types of road damage and has been implemented on two platforms: Jetson Nano and a personal computer or laptop. A comparative analysis was conducted to evaluate accuracy, computational performance, and power efficiency. The study employs YOLO (v5, v7, v8) and EfficientDet algorithms for detecting road damage. Experimental results indicate that EfficientDet achieves the highest accuracy at 88%, while YOLO attains 63%. In terms of computational performance, YOLOv8 delivers the highest frame rate, reaching 25 FPS on the Jetson Nano. Power efficiency analysis reveals that YOLOv8 on the Jetson Nano is six times more energy-efficient compared to its implementation on a laptop. Likewise, EfficientDet on Jetson Nano demonstrates three times better energy efficiency than on a laptop. These findings underscore the feasibility of deploying AI-powered embedded vision systems for detecting road damage. The use of deep learning models on energy-efficient platforms, such as Jetson Nano, enhances real-time performance while minimizing power consumption. Future research should focus on optimizing these models to enhance performance on edge devices while further assessing their practical applications in real-world environments.