GeoSurveyAI technical brief
GeoSurveyAI converts vehicle-mounted video and centimetre-level GNSS telemetry into a queryable, geo-referenced inventory of pavement distress and roadside assets. It uses consumer camera hardware and is checked against LiDAR ground truth. This brief describes how it works, how it is verified, and where its limits are.
The problem
Network-level road condition surveys in India are mostly done either manually, which is slow and subjective, or with dedicated laser-equipped survey vehicles, which are expensive to buy and mobilise. Many authorities therefore survey their networks rarely, or only in part. GeoSurveyAI aims to make a repeatable, evidence-backed survey possible with a camera, an RTK receiver and an ordinary vehicle.
System architecture
1. Capture
A consumer action camera records the road from an ordinary vehicle while an RTK GNSS receiver logs precise position. Video and position are synchronised so that every frame has a known location.
2. Detect
AI models find roadside assets in the forward view and outline pavement distress such as cracking, potholes and rutting. Repeated sightings of the same asset or defect across frames are merged so each is counted once.
3. Locate and measure
Each detection is placed on the map with absolute coordinates. Distress is measured for area, length and estimated depth. Roughness is estimated from the camera's own motion data and the vehicle's speed.
4. Deliver
Results are written as GeoJSON and CSV and shown on an interactive web map with severity-coded distress. The same files can be ingested by the City Level Asset Management System or any GIS that reads standard formats.
Specification summary
| Capture | Consumer action camera and RTK GNSS on a standard vehicle |
|---|---|
| AI analysis | Multi-model detection and segmentation of roadside assets and pavement distress |
| Roughness | IRI estimated from onboard motion data, for network-level screening |
| Standards | IRC:82-2023, IRC:67, IRC:SP:16 |
| Training data | Indian road imagery from public datasets and our own corridor surveys |
| Validation | Terrestrial laser scanning (TLS) ground truth and RTK survey |
| Outputs | GeoJSON, CSV, interactive web map |
How accuracy is verified
- Model selection. Candidate models are compared against terrestrial laser scanning (TLS) data of the same road, using statistical tests, before one is adopted.
- Position accuracy. Asset positions are measured against a manual RTK survey of the same features.
- Benchmarking. Output is compared with commercial road-survey platforms on the same sections.
- Reporting. Results are reported per defect class and per road condition, not as a single headline number, so reviewers can compare them with their own checks.
Validation plan
We are arranging corridor validation with municipal and highway partners in Karnataka, targeting 500 to 1,500 km of network across urban and state roads. Each pilot is scoped with the partner: a representative corridor, their own ground-truth checks, and a report that lists detections against those checks. Partners can start with a single 50 km corridor.
Limits and safeguards
- Results depend on camera mounting, lighting and lens cleanliness; we publish capture guidelines and reject unusable frames.
- IRI from an IMU is intended for network-level screening and prioritisation. Where a tender specifies a particular profiler for acceptance testing, that equipment is still required.
- Models are strongest on the defect and asset classes they were trained on. New scene types need review and, where needed, additional labelled data.
- Every AI flag on a critical asset is confirmed by an engineer before it reaches a report or work order.
Integration
Outputs are standard GeoJSON and CSV, so they load into any GIS or road asset register. They are structured for the City Level Asset Management System, where survey results become condition records on the right road. Ground-truth scans come from our terrestrial LiDAR team.