TinyML-Based Pothole Detection: A Comparative Analysis of YOLO and FOMO Model Performance

Potholes pose a substantial threat in urban settings, leading to vehicular damage and safety risks. Therefore, their prompt identification is essential for effective road maintenance. In this study, we undertake a comparative evaluation of the YOLOv5, YOLOv8, and FOMO models for pothole detection using the Tiny Machine Learning (TinyML) approach. We use an openly available dataset comprising images labeled with bounding boxes, train these models, and gauge their precision in identifying potholes at various input sizes. Furthermore, we track the carbon footprint during the training phase of these networks using Code Carbon. Finally, as we methodically decrease the intricacy of the detection network, we evaluate the effects of this reduction on energy usage and detection efficacy. The outcomes suggest notable performance across all models, with the FOMO model standing out as the quickest and most energy-efficient model for pothole detection.