Detection Pipeline¶
LAIRA detects distresses, markings, and FOD through five primitives that combine into the higher-level detections cited by each matrix. The matrices speak in standards language (D6433 §X1.5 alligator cracking, UFC 3-260-16 PAVER code 41); this page speaks in engineering language (what algorithm runs on what sensor data) and connects the two.
The five primitives¶
Every detection in every matrix can be traced to one or more of these:
| Primitive | What it does | Sensor source | Used for |
|---|---|---|---|
| Crack detection | Identifies linear and pattern cracks; measures width, length, depth, and area | RGB pattern recognition + LiDAR depth | Alligator, block, edge, joint reflection, L&T, longitudinal/transverse, slippage, linear (PCC), shrinkage, durability (D), corner break, divided slab, shattered slab |
| Elevation differential | Detects vertical discontinuities between adjacent surfaces or across joints | LiDAR DEM | Faulting, depression, swell, drop-off, bumps and sags, settlement, blowup, corner spall, joint spall |
| Surface texture | Quantifies micro-texture and material reflectivity changes | LiDAR intensity + RGB color and pattern | Bleeding, polished aggregate, oil spillage, weathering, raveling, jet blast erosion, scaling, map cracking |
| Profile / segmentation | Segments pavement boundaries, lane edges, markings, slab geometry | LiDAR + RGB ML segmentation | Edge cracking, lane/shoulder drop-off, marking classification, PCC slab counting, FOD object isolation |
| Zone-context classification | Triggers distress evaluation based on contextual zone (e.g., railroad track, joint, hold-short bar) | RGB scene-context model | Railroad crossing distress (D6433 X1.18 / X2.17), hold-short marking compliance, taxiway-vs-runway evaluation |
Crack pipeline detail¶
The crack pipeline is the workhorse — most matrix rows in the AC distress sections trace through it.
RGB frame ──▶ Crack-detection CNN ──▶ Crack mask (binary)
│ │
│ ▼
│ Crack centerline + skeleton
│ │
▼ ▼
LiDAR depth ──▶ Per-pixel depth ──▶ Crack width + depth + length
│
▼
Severity classification
(per standard's L/M/H tables)
│
▼
Per-detection record
(type, severity, geolocation,
timestamp, image)
The output is structured: each crack carries its standards-mapped distress type, severity per the relevant standard's threshold table, geolocation at ±2 cm, timestamp, and reference imagery. This is what populates every "Per-detection record" output cited in the matrices.
Elevation-differential pipeline detail¶
Every elevation-based distress runs through the same primitive. The variation across standards is in the threshold tables (D5340 vs D6433 vs PROWAG), not the underlying measurement.
LiDAR point cloud ──▶ Surface DEM (gridded) ──▶ Local elevation field
│
▼
Differential against reference
(joint, edge, adjacent panel,
longitudinal baseline)
│
▼
Severity per threshold table
(D6433 §X2.7, PROWAG R302.7.2,
etc.)
│
▼
Per-detection record
Resolution: ±3 mm at typical operating heights. Below the smallest tabulated severity threshold in any pavement standard.
Marking detection¶
ML segmentation classifies each marking by type (runway centerline, taxiway edge, lane line, crosswalk, hold-short, stop bar, arrows, surface-painted signs) and runs three subordinate checks:
- Width and pattern — segmentation mask × LiDAR depth measurement.
- Color compliance — RGB color classification against the relevant standard's color spec.
- Retroreflectivity (screening) — calibrated LiDAR reflectivity, used as a screening estimate, never as a replacement for an ASTM E1710 handheld retroreflectometer reading.
The marking model has both an airfield class library (AC 150/5340-1M, UFC 3-260-04) and a roadway class library (MUTCD 11th Ed., Part 3). They share the same backbone segmentation network with different classification heads.
FOD detection¶
FOD detection runs on every frame at 10 Hz with cross-frame deduplication:
| Step | What it does |
|---|---|
| Object proposal | RGB + LiDAR height filter generates candidates above ~1 cm |
| Material classification | LiDAR reflectivity ratio + RGB classifies into Metal / Stone / Plastic / Rock / Rubber |
| Geolocation | RTK fix per detection (±2 cm) |
| Deduplication | Cross-frame matching prevents the same object being counted multiple times within a pass |
| Severity scoring | Configurable threshold for false-alarm-rate management per AC 150/5220-24 |
How primitives compose into matrix rows¶
A single matrix row often combines two or three primitives. Example: PCC corner break (D6433 §X2.4) requires both the crack-detection primitive (to identify the crack) and the profile/segmentation primitive (to confirm the crack intersects a slab corner per §X2.4 geometry). Joint spall (D6433 §X2.21) combines the elevation-differential primitive (depth of material loss) with surface texture (boundary of the spall against intact pavement).
This composition is why a single sensor upgrade or algorithm improvement tends to flip multiple matrix rows at once — the primitives are shared across many distress types.
Where matrices cite this page¶
| Matrix | Rows that anchor here |
|---|---|
| Civil Airfields | Pavement Condition Assessment, FOD Detection, Pavement Markings sections |
| Military Airfields | §1 Asphalt distresses, §2 Concrete distresses, §4 FOD, §5 Marking Detection |
| Federal & State Roads | §1.3 AC distresses, §1.4 PCC distresses |
| Roadway Markings (MUTCD) | Marking classification + retroreflectivity screening |
| Sidewalks & Curb Ramps | PROWAG geometric primitives (cross slope, running slope, vertical discontinuity) |