RoadScout
RoadScout platform · Full capability map

Complete road intelligence

Five production vision models, 25 Hz motion telemetry, scoring to the standards your engineers already use, and a platform that carries every finding through review, work order and forecast. Here is the full apparatus, named and measured.

RoadScout Vision

Five models. 65+ classes. 74,000+ training images.

Every frame from every trip passes through five production YOLO models. We publish the class lists and the measured accuracy, because a class you cannot name is a defect you cannot budget for.

SEC. 01 / COMPUTER VISION

Surface distress

26 CLASSES
RDD2022 MODEL

The core defect model, trained on the RDD2022 dataset and extended for local conditions. Every detection carries a confidence score and severity grade.

PotholeCrack: longitudinalCrack: transverseCrack: alligatorCrack: blockCrack: wheel-trackRuttingRavelling (mod/sev)Chipping lossSubsidenceEdge deteriorationBinder bleedingDefective overlayPatch: areaPatch: spotSeal: longitudinalSeal: transverseSurface treatments ×4Traffic signSurface markingManholeStorm drain

Traffic signs

29 CLASSES
SIGNS MODEL · mAP50 0.677 · MEASURED

Every sign is detected, classified and given a 1 to 5 condition rating, so the inventory records not just where signs are but what state they are in.

StopGive waySpeed limits 10 to 120No entryNo parkingParkingNo overtakingPedestrianCrosswalkRoundaboutTurn left / rightKeep left / rightSignal heads R/Y/GBumpy road

Road markings

14 CLASSES
MARKINGS MODEL · mAP50 0.942 · BEST MODEL

Our most accurate model. Line and surface markings are catalogued on every pass, and telemetry flags stretches where lane lines are missing entirely.

Lane lineCenter lineBoundary lineStop lineCrosswalkArrows ×5Junction boxBicycle laneBus lane

Infrastructure

8 CLASSES + COCO ASSETS
INFRASTRUCTURE MODEL · mAP50 0.711 · MEASURED

Roadside assets are picked up alongside the surface, so the asset register updates itself on every trip. COCO-based detection adds common street furniture.

Street lightGuardrailManholeDrainSpeed bumpRumble stripUtility poleBollardTraffic lightStop signFire hydrantBenchParking meter
SEVERITY GRADING

Three severity bands, by confidence

HIGHconfidence ≥ 0.60
MEDIUMconfidence ≥ 0.45
LOWbelow 0.45
10 M CLUSTERING

One defect, one event

The same pothole seen in twelve consecutive frames is one pothole. Detections are clustered within a 10 m radius, with cross-model deduplication, so the map shows real defects, not frame counts.

EVIDENCE FRAMES

The top 3 frames, kept

Every clustered event keeps its three best evidence frames, GPS fix and confidence score: proof for the review decision, the work order, the council meeting or the contractor dispute.

Telemetry intelligence

What the camera cannot see, the motion sensors feel.

Alongside the frames, RoadScout streams 25 Hz accelerometer, GPS and altitude data. Four analyzers turn that motion signature into findings no vision model can produce.

SEC. 02 / 25 HZ IMU

CORRUGATION25 HZ · 5 S WINDOWS

Corrugation detection

The accelerometer signal is analysed in 5-second windows to flag corrugated sections, the washboard surfaces that vision alone underreports, especially on gravel and haul roads.

GRADE≥6% OVER ≥50 M

Grade and inclination

GPS altitude identifies sustained grades of 6% or more over at least 50 m, tagged uphill or downhill: input for drainage, safety and haul-route planning.

LANE LINESOVER ≥100 M

Missing lane lines

Where the markings model finds no lane lines over a continuous 100 m or more, the stretch is flagged as a missing-marking event for the repainting programme.

ROUGHNESSIRI ESTIMATE

Roughness to IRI

Vertical acceleration is converted to an IRI estimate in m/km, banded to World Bank / ARRB thresholds. An estimate from 25 Hz telemetry, stated as such, not a quarter-car measurement.

Scoring & standards

Scored to the scales auditors accept.

Detections become decisions only when they are scored on scales engineers, auditors and funders recognise. RoadScout reports PCI, IRI, TMH and a composite Trip Score for every trip, per roughly 100 m segment.

SEC. 03 / STANDARDS

PCI · ASTM D6433 · SEVERITY-WEIGHTED DEDUCT MODEL · PER ~100 M SEGMENTFIG. 01
GOOD 86+SATISFACTORYFAIRPOORVERY POORSERIOUSFAILED

Detected distresses feed a severity-weighted deduct model to produce a Pavement Condition Index on the ASTM D6433 seven-band scale. The good-practice benchmark the platform tracks against is PCI 85.

IRI · WORLD BANK / ARRB BANDSFIG. 02
Excellent< 1.5 m/km
Good< 2.7 m/km
Fair< 4.2 m/km
Poor< 6.0 m/km
Very Poor≥ 6.0 m/km

Roughness in m/km, estimated from 25 Hz telemetry and banded to World Bank / ARRB thresholds. Good-practice benchmark: IRI 2.7. Stated plainly: this is an estimate, not a quarter-car profilometer reading.

TRIP SCORE · 0 TO 100 · COMPOSITEFIG. 03
Surface 40%Signage 20%Markings 20%10%10%
Surface40%
Signage20%
Markings20%
Infrastructure10%
Severity10%

One number per trip, weighted across surface, signage, markings, infrastructure and severity, with road-type context: highway, urban, suburban or rural.

TMH · VISUAL CONDITION · SA / SADC

TMH 5-point visual condition, built in

South African and SADC authorities report against the TMH 5-point visual condition scale. RoadScout generates it automatically alongside PCI and IRI, so regional submissions need no manual re-scoring.

SURFACE TYPES · 9 CLASSIFICATIONS

Sealed, unsealed and everything between

Asphalt concreteAsphalt overlaySurface treatmentConcreteBlock pavingCobblestoneGravelEarthPaved-gravel mixed

Dashboard & platform

From map to work order to next year's budget.

Detection is the start, not the product. The RoadScout dashboard carries every finding through review, work order, verification and forecast, and hands the data back out through GIS exports and a read-only API.

SEC. 04 / PLATFORM

MAP

Interactive network map

Every clustered event on a map of your network, with the top 3 evidence frames, GPS fix, confidence score and severity grade attached. Filter by type, severity, status and date.

REVIEW

Accept / decline workflow

Engineers review detections from the desk: accept, decline, or reprioritise, each decision logged with an audit trail. The record shows who confirmed what, and when.

WORK ORDERS

From detection to job card

Raise work orders straight from accepted events, with priority, status, assignee and estimated cost. Track them from open to verified without leaving the platform.

INSPECTIONS

Virtual inspections

Walk a route from the desk using the captured frames and scores instead of driving out. Time-stamped visual records support safety audits and cut repeat site visits.

REPORTS

Reports and inventories

Trip summaries, defect inventories, sign inventories with condition ratings, and marking inventories, generated from the data your fleet already collected.

FORECAST

Network analytics and deterioration forecast

PCI and IRI across the network, condition distribution, benchmark views, and a deterioration forecast that turns today's scores into a multi-year works programme.

GIS

GIS exports and API

Shapefile, GeoJSON, KML and CSV in WGS84, straight into ArcGIS, QGIS or your PMS. A read-only REST API and webhooks keep your own systems the source of record.

RETRAINING

Human-in-the-loop retraining

Review decisions feed training annotations, jobs and metrics, so the models keep improving on your network's roads, surfaces and signage.

See every feature on your own roads.

A demo takes one call. A pilot takes one phone, one vehicle and a week of normal rounds.