Performance Evaluation of YOLO26 for Pothole Detection Under Diverse Environmental Conditions

Abstract

Road potholes are a serious threat to road safety. They can damage vehicles and increase the risk of traffic accidents. However, factors like lighting and weather can impact how well computer vision systems detect potholes. To tackle this issue, this study compares four You Only Look Once (YOLO) 2026 (YOLO26) variants: YOLO26n, YOLO26s, YOLO26m, and YOLO26l, for detecting potholes in different environmental conditions. A dataset was independently collected in the Poris area of Tangerang, Indonesia, using a forward-facing camera to capture a driver's view. The video recordings were turned into annotated image frames in YOLO format, creating 368 images that included clear, rain, evening, and night conditions. Each model was trained for 200 epochs under the same experimental conditions. Their performance was evaluated using Precision, Recall, mean Average Precision at an Intersection over Union (IoU) threshold of 0.5 (mAP@0.5), and mean Average Precision across IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95). Computational efficiency was evaluated using parameter count, model size, training time, GPU memory usage, inference time, and frames per second (FPS) to assess the computational requirements and real-time detection capability of each YOLO26 variant. The results showed that YOLO26m achieving the highest Recall of 0.877 and mAP@0.5 of 0.934, while YOLO26l achieved the highest Precision of 0.989 and mAP@0.5:0.95 of 0.735. On the other hand, YOLO26n was the most computationally efficient option while still providing competitive detection performance. These results highlight the trade-off between detection accuracy and computational efficiency among YOLO26 variants. They also offer useful guidance for choosing the right model for real-time pothole detection in various conditions, which can help in the development of future driver assistance systems and pothole warning systems.