How Do New Digital Eyes Detect Potholes and Cracks on Roads?

Roads need constant care, but spotting potholes and cracks isn’t always simple. Researchers have built TriRoad AI, a web platform that uses artificial intelligence to detect pavement defects in real time. The core of the system are two YOLO (You Only Look Once) versions, YOLOv8 and YOLOv11, known for fast object recognition. Users upload a road photo from their browser: the algorithm flags damaged spots, assigns a severity level, and places them on an OpenStreetMap.

Tests were run on two datasets: images downloaded from the internet and field photos taken in Ankara, Turkey. YOLOv8 achieved 88.43% accuracy on online data, while YOLOv11‑B proved more robust on the ground, reaching 46.15% on mixed images. The Gemini model shone in the lab (97.64% accuracy) but dropped to 18.50% in real‑world conditions.

TriRoad AI stands out because it needs no installation—just open the site and upload pictures. It also merges damage detection with an interactive geospatial map, easing preventive maintenance planning. This approach could make cities smarter, cut costs, and boost road safety.

The YOLO engine belongs to the single‑stage detector family, processing the whole image in one pass, unlike two‑stage methods (e.g., R‑CNN) that first propose candidate regions. YOLOv8 introduces an enhanced CSPDarknet backbone and a decoupled regression head, allowing better separation between classification and localization. YOLOv11‑B, built on a transformer‑lite architecture, leverages global attention to capture broader context, useful when cracks are thin or partially occluded.

The Ankara dataset was manually annotated with bounding boxes and severity levels (light, medium, severe). To assess robustness, a confidence filter was applied: only predictions above 0.6 are displayed, reducing false positives. The OpenStreetMap layer was enriched with GeoJSON containing GPS coordinates of damages, enabling spatial queries (e.g., "show all damages within 500 m of a point").

Gemini, Google's multimodal model, combines vision and language. In controlled settings, its ability to learn from massive datasets yields excellent performance. However, its architecture requires high‑quality inputs and a data distribution similar to training; in urban scenarios with lighting variations, angles, and obstacles, generalization drops sharply. This underscores the need to test AI models under real‑world conditions before large‑scale deployment.