Application of Crack Identification Techniques for an Aging Concrete Bridge Inspection Using an Unmanned Aerial Vehicle

Tests a pipeline on an ageing concrete bridge that detects cracks in photos from a commercial drone and measures their width and length. Cracks were detected with an R-CNN fine-tuned on a small training set, and a 70 mm square marker attached to the structure gave the pixel scale, with homography correcting distortion. Flights were manual at about 2 m because GPS was unreliable under the bridge. Twelve field cracks were sized, but without ground-truth checks on site.

Key findings

• Lab test relative error about 1-2% for crack quantification
• Field: 12 cracks measured at about 0.53-2.47 mm thick and 6.6-78.4 mm long
• Processing about 150 min for the point cloud and about 14 min for deep-learning detection

Limitations

• Field measurements not checked against ground truth; no precision/recall reported
• Requires a physical marker on the structure for scale
• Manual flight; GPS unreliable under the bridge
• One bridge, only the rails inspected; small training set (384 images)