3D Crack Localization via AI and Multi-View Reconstruction
Manual inspection of structures is subjective and difficult to quantify, while traditional 2D crack detection methods lack spatial context. This project proposes a 3D inspection pipeline that combines AI-based crack detection with geometric reconstruction to enable spatial localization and analysis of surface defects.
- Detect surface cracks using deep learning
- Reconstruct inspected surfaces in 3D
- Localize cracks directly in 3D space
- Fuse evidence from multiple views
- Robustly fuse evidence from multiple views
3D point cloud with crack localization obtained by projecting 2D crack predictions into the reconstructed geometry.
Multi-view 3D point cloud with crack localization obtained by projecting 2D crack predictions into the reconstructed geometry.
{
"scene_status": "DAMAGED",
"num_detected_cracks": 6,
"inspection_parameters": {
"confidence_threshold": 0.6,
"min_crack_length": 1.0,
"min_cluster_points": 10
},
"cracks": [
{
"id": 0,
"length": 3.66,
"num_points": 213,
"center_3d": [
10.057,
5.434,
15.151
]
},
{
"id": 1,
"length": 2.26,
"num_points": 76,
"center_3d": [
3.055,
5.303,
15.335
]
},
....
- Reconstruction: COLMAP-based multi-view 3D reconstruction
- Detection: U-Net-based crack segmentation on 2D images
- Fusion: Projection of 2D predictions into 3D space and multi-view aggregation
- Reporting A generated report .json showing the cracks details
- Multi-view images are captured
- COLMAP reconstructs geometry and camera poses
- Crack detection is applied per image using a pretrained AI model
- Predictions are projected into 3D
- Multi-view fusion builds a 3D crack heatmap
- Reporting
- Offline inspection (not real-time)
- Requires sufficient image overlap
- Detection accuracy depends on training data