A road-audit console built from ordinary dashcam video
Three AI modules turn every drive-through into a classified, quantified, GPS-tagged inventory of broken surfaces, roadside assets, and signboards — with decision-ready reporting at the end.
- 3 AI modules
- 34,540+ training images
- IRC · MoRTH · NHAI aligned
Three AI modules on every frame
Three models run on every frame. Pavement reads the running surface — potholes, cracking, kerb condition, each measured and graded. Infrastructure inventories what sits beside it, from lane lines to crash barriers. Signage turns every board into a typed, located asset record of its own.
Pavement Intelligence
The pavement module reads every frame for damage — and grades what it finds by severity and measured dimensions, rather than just flagging it.
- Potholes
- Cracks
- Kerb condition
- Severity & measurements
- Repair recommendations
Infrastructure Intelligence
The infrastructure module outlines the road furniture along the route in clear polygons, so a missing barrier or faded kerb paint shows up in the ledger.
- Lane & edge lines
- Kerb paint
- Road barriers
- Signboards & road assets
- Raised pavement markers
- Rumble markings & speed breakers
Signage Intelligence
The signage module files one record per physical board — its category, whether it still reflects at night, and exactly where it stands.
- Informative / Regulatory / Warning
- Reflective or non-reflective
- Signboard annotation
- Location-based evidence
A YOLOv8/v11 ensemble, trained on Indian roads
Trained on 34,540+ augmented images and processing video at 45–60 FPS, the stack finds, labels, and pins every defect it passes — with monocular depth estimation (MiDaS) supporting real-world measurements — while the vehicle drives at normal traffic speed.
Five outputs from every run
A single run produces five deliverables: annotated media of each detection, a table carrying class, confidence, measurements and coordinates, an interactive map of the corridor, the processed video, and a CSV export. Five forms of the same truth — explorable by engineers, filable by auditors.
Annotated media
Clear visual detection evidence — every defect and asset outlined on the original frame.
Detection tables
Class, confidence, measurements and GPS for every finding — searchable and filterable.
Interactive maps
Street or satellite view with colour-coded markers, routes, and tap-to-view details.
Processed videos
Review annotated footage with play, pause, seek and fullscreen — no extra software.
Data export
Download structured CSV results ready for spreadsheets, audits, and asset systems.
Twelve things the platform does for you
Detection is one capability of twelve. The rest are the working platform around it: secure team access, capture from image, video or live feed, a processed-video player, explorable maps, reports, a city-wide overview, repair recommendations with cost estimates, reliability monitoring, and desktop plus mobile.
Secure team access
Each team member logs in securely — field staff and directors each see the dashboard that suits their job.
Capture data your way
Upload a photo, upload a video, or use a live camera. GPS location is captured automatically.
Pavement problem detection
Automatically finds potholes, cracks and damaged kerbs, and reports their size and severity.
Infrastructure detection
Spots lane markings, kerb paint, signboards and road barriers, all clearly outlined on the image.
Signage detection
Identifies each signboard's type and whether it's reflective, with one clear record per sign.
Watch processed video
Review annotated videos with normal play, pause, seek and fullscreen controls.
Maps you can explore
Every finding on a street or satellite map, colour-coded, with tap-to-view details and routes.
Ready-to-use reports
Results organised into clear tables you can search, filter, and export to a spreadsheet.
City-wide overview
Summary charts and stats across all roads, dates and inspectors — the bigger picture.
Repair recommendations
Suggested repairs, priority levels and cost-estimate support for each defect found.
Always reliable
The AI system is continuously monitored behind the scenes, so results stay accurate.
Works everywhere
A clean, consistent design that works smoothly on desktop and mobile.
Run like production software, not a science project
Behind the detections sits a full MLOps pipeline: versioned datasets, automated data validation, drift monitoring, a model registry, lineage tracking, and live metrics. It exists so a result from this month is reproducible next year — which is what makes an audit defensible rather than merely recent.
Versioned datasets
Every training dataset is hash-versioned, so any model can be traced to the exact data it learned from.
Validated before training
Automated data-quality gates check every image and label before they ever reach a model.
Drift watched
Incoming data is compared against training data, so quality shifts are caught before they hurt accuracy.
Every experiment tracked
Runs, metrics, and artifacts live in a model registry — models are promoted through review, never overwritten.
Full lineage
A recorded graph links raw images to the exact model version behind every detection.
Monitored round the clock
Live health, latency, and accuracy dashboards with automatic alerting when anything degrades.
Every finding is traceable to a location, a frame and a confidence score — and every model to the data that trained it.
Manual surveys vs YNMDrishti
A manual survey takes weeks of lane-by-lane inspection and produces subjective judgements on paper, outdated by the time they are filed. One drive-through at traffic speed produces quantified severity, affected area, and GPS-tagged exports — typically the same day.
Manual surveys
- Weeks of lane-by-lane inspection
- Subjective visual judgement
- Findings outdated by the time they're filed
- Paper trails no auditor can follow
With YNMDrishti
- One drive-through at normal traffic speed
- Quantified severity and affected area
- Same-day results, typical
- GPS-tagged, version-controlled exports
Speaks the language your auditors audit in
Defect classes and severity grades map to Indian Roads Congress, MoRTH and NHAI categories at detection time, not afterwards. A deliverable that already uses the categories a submission expects goes straight in — a vendor taxonomy has to be translated first, and translation is where defensibility leaks away.
IRC alignment
Severity grading uses the vocabulary Indian Roads Congress auditors already work in.
MoRTH alignment
Report structure lines up with Ministry of Road Transport & Highways formats.
NHAI alignment
Kilometre-wise inventories drop straight into NHAI review packs.
Data handling
Your footage and findings stay yours — processed on your terms, never shared.
What evaluators ask us about the product
What road defects can YNMDrishti detect?
Three modules cover the road between them. Pavement finds potholes, cracking and kerb condition with severity and measurements. Infrastructure covers lane and edge lines, kerb paint, barriers, roadside assets, road studs, rumble markings and speed breakers. Signage records each sign's type and reflectivity as its own asset.
What do I actually receive at the end of a run?
Five things: annotated media showing each detection, a table carrying class, confidence, measurements and coordinates, an interactive map of the corridor, the processed video, and a CSV export. The CSV matters most in practice — it is what lets findings enter whatever system you already run.
Do I need special hardware or a survey vehicle?
No. A standard dashcam recording 1080p at 30 fps, mounted in a vehicle already making the trip, is the whole capture requirement. There is no laser profiler, no sensor mast, and no lane closure, because the vehicle travels at ordinary traffic speed.
Does the output map to IRC, MoRTH and NHAI categories?
Yes — defect classes and severity grades are aligned to those categories, which is what lets a deliverable go into a submission without being translated first. Translation is where defensibility usually leaks away, so the mapping is done at detection time rather than afterwards.
How accurate is the detection?
85% on pothole detection and 97% on road-surface classification, from an ensemble trained on 34,540+ augmented images of Indian roads. Those measure different tasks and should not be averaged. Every detection carries a confidence score and links to its source frame, so any row can be checked.
Put YNMDrishti to work on your roads
From dashcam video to a regulator-ready audit, typically in a day. Send us a stretch of your network and we'll run the first pass.
- 85%pothole detection
- 34,540+training images
- IRC · MoRTH · NHAIaligned reporting
- Same daytypical turnaround
No specialist hardware, no lane closures — and a complete sample audit you can open right now without talking to anyone.