Robust Pothole Detection Using YOLOv5 and SSD: Accuracy, Speed, and Deployment Trade-offs

This study presents a comparative evaluation of two state-of-the-art object detection algorithms, YOLOv5 and SSD (Single Shot MultiBox Detector), for the task of pothole detection on urban and rural road surfaces. Both models were trained and tested on a curated and annotated dataset encompassing diverse lighting conditions, occlusions, and road environments. Experimental results demonstrate that YOLOv5 outperforms SSD across key performance metrics including Precision (0.914 vs. 0.846), Recall (0.872 vs. 0.803), F1-score (0.892 vs. 0.824), and mean Average Precision ([email protected]: 0.918 vs. 0.847). YOLOv5 also exhibited greater robustness in detecting small and partially occluded potholes, and achieved better generalization when evaluated on an external dataset. Despite YOLOv5’s slightly slower inference speed (78 FPS vs. 108 FPS for SSD), its improved localization accuracy and lower false positive rate make it more suitable for safety-critical deployments such as autonomous vehicles and road maintenance systems. SSD, while less accurate, demonstrates superior speed and efficiency, making it a favorable option for embedded or latency-sensitive applications. The findings suggest that the selection between these models should be guided by the intended deployment context, balancing the tradeoffs between accuracy and computational efficiency. This work highlights the importance of robust detection under challenging real-world conditions and points to future directions for improving model performance through dataset enrichment and hybrid architectural approaches.