Bridges are essential components of transport infrastructure, maintaining regional accessibility and public safety, yet they face constant deterioration from aging materials, environmental exposure, and mechanical stress. The 2018 Polcevera Viaduct collapse led Italy's Ministry of Infrastructure to establish national Bridge Inspection Guidelines, creating standardized risk assessment protocols, using a multi-level approach that relies on visual inspections and defect classification through severity (G) parameters, extension (K1), and intensity (K2). Current inspection practices depend heavily on manual assessment, proving time consuming, costly, and inconsistency due to subjective defect identification. This study proposes a semi-automated approach combining drone imaging with deep learning and 3D modeling to streamline bridge inspections. The method employes UAV images to detect surface defects in bridge structures through YOLO11, a real-time detection model. After defect detection, the Segment Anything Model (SAM2) is applied to create precise defect masks and bridge component segmentation, enabling accurate geometric measurements and spatial mapping. The detected defects are located within a 3D bridge model reconstructed using Structure-from-Motion photogrammetry, ensuring real-world coordinate accuracy. The K1 and K2 indices, as required by Italian inspection standards, are automatically derived. The methodology is validated on a reinforced concrete overpass in Lombardy and compared with traditional field inspection results. Findings show this workflow provides a consistent and effective alternative to manual inspection while maintaining required accuracy. The methodology enables scalable bridge assessment across complex infrastructure networks.
I ponti rappresentano elementi fondamentali dell'infrastruttura di trasporto. Tuttavia, essi sono soggetti a continuo deterioramento dovuto all'invecchiamento dei materiali, esposizione ambientale e sollecitazioni meccaniche. Il crollo del Viadotto Polcevera nel 2018 ha portato il Ministero delle Infrastrutture e dei Trasporti a istituire le Linee Guida nazionali per l'Ispezione dei Ponti con protocolli multi-livello standardizzati basati su ispezioni visive e classificazione dei difetti secondo parametri di gravità (G), estensione (K1) e intensità (K2). Le attuali pratiche ispettive dipendono dalla valutazione manuale, risultando dispendiose e soggettive nell'identificazione dei difetti. Questo studio propone un approccio semi-automatizzato che combina acquisizione di immagini mediante droni con deep learning e modellazione 3D. Il metodo utilizza immagini scattate da droni per rilevare difetti superficiali attraverso il sistema di rilevamento di difetti in tempo reale YOLO11. Successivamente, il modello SAM2 crea maschere precise dei difetti e segmentazione dei componenti per misurazioni geometriche accurate. I difetti rilevati vengono localizzati in un modello 3D ricostruito mediante fotogrammetria con il processo Structure-from-Motion, derivando automaticamente gli indici K1 e K2 richiesti dagli standard italiani. La metodologia è validata e confrontata con un’ispezione tradizionale su un cavalcavia in cemento armato in Lombardia. I risultati mostrano che questo approccio fornisce un'alternativa coerente ed efficace all'ispezione manuale mantenendo la precisione richiesta, consentendo valutazioni scalabili attraverso reti infrastrutturali complesse.
Semi-automated bridge inspection using drones and deep learning
Khosravi, Pouria;Jalali Varnamkhasti, Zahra
2024/2025
Abstract
Bridges are essential components of transport infrastructure, maintaining regional accessibility and public safety, yet they face constant deterioration from aging materials, environmental exposure, and mechanical stress. The 2018 Polcevera Viaduct collapse led Italy's Ministry of Infrastructure to establish national Bridge Inspection Guidelines, creating standardized risk assessment protocols, using a multi-level approach that relies on visual inspections and defect classification through severity (G) parameters, extension (K1), and intensity (K2). Current inspection practices depend heavily on manual assessment, proving time consuming, costly, and inconsistency due to subjective defect identification. This study proposes a semi-automated approach combining drone imaging with deep learning and 3D modeling to streamline bridge inspections. The method employes UAV images to detect surface defects in bridge structures through YOLO11, a real-time detection model. After defect detection, the Segment Anything Model (SAM2) is applied to create precise defect masks and bridge component segmentation, enabling accurate geometric measurements and spatial mapping. The detected defects are located within a 3D bridge model reconstructed using Structure-from-Motion photogrammetry, ensuring real-world coordinate accuracy. The K1 and K2 indices, as required by Italian inspection standards, are automatically derived. The methodology is validated on a reinforced concrete overpass in Lombardy and compared with traditional field inspection results. Findings show this workflow provides a consistent and effective alternative to manual inspection while maintaining required accuracy. The methodology enables scalable bridge assessment across complex infrastructure networks.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_07_Jalali_Khosravi.pdf
solo utenti autorizzati a partire dal 01/07/2029
Descrizione: Text of the thesis
Dimensione
12.5 MB
Formato
Adobe PDF
|
12.5 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/240866