Real Time Deep Learning Based Pothole Detection on Low Cost Embedded Devices for ADAS

Potholes and road damages compromise public safety and vehicle longevity by increasing accident risks and maintenance costs. This study evaluates three object detection models—YOLO, RT-DETR, and RetinaNet—for automated pothole detection in ADAS. A dataset of 3777 annotated images, captured from a vehicle’s perspective, was utilized with techniques to address class imbalance and enhance robustness. Among the evaluated models, YOLOv11 achieved the highest mAP50 of 72.3% at 640x640 resolution, while maintaining low inference times on RTX 3060 and Raspberry Pi. These results highlight YOLO’s superior performance in accuracy and efficiency, demonstrating its viability for real-time pothole detection in ADAS applications, even on resource-constrained hardware.