AI detection and road asset inventory
AI road detection and infrastructure inventory
One drive down a street produces tens of gigabytes of imagery and point cloud data. But a city does not need recordings, it needs answers: where the signs are, which of them are damaged, how many potholes have appeared since last year. That layer is what the AKYS AI models build. They read every frame, recognise the objects in it, and turn them into a queryable road asset inventory in which each finding carries its own GPS coordinates.
Why 11K 360° resolution is what makes small-object detection possible
No AI model can recognise something that is not in the pixels. This is exactly where ordinary street photography breaks down: a road sign standing fifty metres ahead occupies a few dozen pixels, and a hairline crack in the asphalt vanishes into image compression altogether. When a model misses that object, it is not because it was trained badly; it is because there was nothing there to analyse.
The Insta360 Titan captures 11K 360° imagery with 10-bit colour depth, and that difference in detail is a practical one. The model reads the face of a sign and identifies its type, instead of concluding only that there is "a pole with a plate on it". It separates road marking that is still crisp from marking that has begun to fade. It notices the thin surface lines that a Baltic winter will turn into potholes. The 10-bit colour helps precisely where contrast is poor: in shadow, at dusk, on worn asphalt.
The imagery does not work alone. It is hardware-synchronised with dual high-density mobile LiDAR and two ZED 2i stereo cameras, so every detection gains geometry as well: real-world distance, height and dimensions.
11K 360°, 10-bit colour
Insta360 Titan. Enough detail for a sign face, fading markings and a thin crack.
Dual LiDAR + stereo
Front near-360° and rear tilted LiDAR, plus 2 × ZED 2i for depth and dimensions.
±3 cm on every finding
RTK-GNSS and a 6-axis IMU. Each detection gets accurate coordinates immediately.
Thermal and surround vision
LWIR thermal imaging and 5 × fisheye cameras for full environmental context.
What the models detect and catalogue
Detection happens in layers. The first finds individual objects, the second segments the scene, the third tracks movement through it. Together they produce not a list but a picture of the city that can be interrogated.
| Road signs | Road sign detection and cataloguing with GPS coordinates |
|---|---|
| Surface damage | Pothole detection, cracks and asphalt damage |
| Lanes | Lane detection and lane geometry, the basis of the road lane defect audit |
| Road markings | Faded, worn or missing markings; crosswalks |
| Vertical assets | Poles and street lights |
| Vegetation | Growth that poses a risk to traffic or sightlines |
| Semantic segmentation | Roads, pavements, buildings, vegetation |
| Motion tracking | Spatial tracking of vehicles, pedestrians, cyclists and traffic flow |
| 3D depth | AI 3D depth reconstruction for real-world distances |
Lane detection and the road marking audit
Faded road marking is one of those defects every driver notices and nobody records systematically. From behind the wheel the line looks "still visible", until it rains, or the light goes. That is exactly the moment lane geometry starts to matter, both to the driver and to any autonomous system reading the road.
The AKYS models recover lane geometry and grade marking condition across the entire route: where a line has faded, where it is worn, where it has disappeared altogether, where a crosswalk no longer matches its own outline. The output is not a photo album but a road lane defect audit: a list of segments with coordinates that can be sorted by condition and used directly as a repainting plan.
Lane geometry
The models reconstruct each lane line and its course across the driven route.
Faded marking grading
Worn and missing markings are graded; 11K resolution is what makes the difference visible.
Segments, not frames
Defects are delivered as road segments with coordinates rather than isolated images.
A sortable work queue
Sort the audit by condition and use it as a repainting plan for the season.
From detection to an inventory with coordinates
A detection on its own is worth little. The value appears when the finding knows where it is. RTK-GNSS combined with a 6-axis IMU delivers ±3 cm geospatial accuracy, so every sign or pothole the model spots becomes a record with an exact location, an object type and a link back to the imagery it came from.
An inventory like that already answers practical questions: show every prohibitory sign in this district, pull the potholes on the streets scheduled for resurfacing, compare marking condition against last year's drive. We deliver the results as GeoJSON or CSV, ready to load straight into a municipal GIS, or to explore in a browser-based digital twin.
GeoJSON / CSV inventory
A filterable asset list with coordinates, prepared for GIS systems.
11K 360° panoramas
Visual evidence behind every record: look at the object yourself.
Point cloud (.las / .laz)
Geometry for measurement, design work and BIM environments.
Digital twin
A browser environment where the inventory and the imagery sit together.
Custom AI models trained for your city
The base AKYS models are trained on European road assets; they understand the signs, markings and surface types actually used here. But every city has its own specifics: local barriers, cycle lane treatments, particular lighting or bus shelter designs, sometimes heritage elements that exist nowhere else.
For those cases we train the models further, on data captured in your own city. The same infrastructure feeds the opposite direction too: AKYS produces COCO-formatted multisensor AI training datasets covering LiDAR point clouds, 11K 360° imagery and thermal data, available through marketplace.akys.ai. If you want to talk through your city's asset list, write to [email protected].
GDPR: anonymisation before storage
A street is full of people and vehicles, and a city dataset must never become a surveillance tool. Anonymisation in the AKYS pipeline therefore happens at the start rather than at the end: dedicated AI models find and blur faces and vehicle licence plates before the data reaches storage.
This costs nothing in analytical terms: signs, surfaces, poles and markings stay fully visible. Only what constitutes personal data is removed, in line with EU GDPR requirements.
Frequently asked questions
What do the AI models detect automatically?
The models find and catalogue road signs, potholes, cracks and other asphalt damage, faded or worn road markings, poles, crosswalks, street lights and vegetation that poses a risk to traffic or visibility. A semantic segmentation layer separates roadway, pavements, buildings and vegetation, while spatial motion tracking analyses vehicles, pedestrians, cyclists and traffic flow. Every catalogued object carries GPS coordinates.
Why does 11K 360° resolution matter for AI detection?
A model can only recognise what is physically present in the pixels. In ordinary street imagery a sign fifty metres ahead or a hairline crack in the asphalt occupies only a few pixels, so it simply disappears before the model ever sees it. The Insta360 Titan captures 11K 360° imagery in 10-bit colour, which leaves enough detail for the models to read a sign face, tell fresh road marking apart from marking that has started to fade, and pick up thin surface cracks before they open into potholes.
Do you detect lanes and faded road markings?
Yes. The lane detection model recovers lane geometry across the whole driven route, and marking condition grading shows where a line is faded, worn or missing entirely. The result is a road lane defect audit: a list of road segments with coordinates that you can sort by condition and use directly as a repainting plan, rather than a folder of photographs.
How do detections become a usable road asset inventory?
Every detection is tied to an accurate position by RTK-GNSS combined with a 6-axis IMU, working to ±3 cm. A finding is therefore not a mark on a photo but a record with coordinates, an object type and a link back to the imagery and point cloud. We deliver the whole set as GeoJSON or CSV, so you can filter it by object type, street or district and load it straight into your own GIS.
Can you train custom AI models for our city's own assets?
Yes. Our base models are trained on European road assets, but every city has its own elements: local barriers, cycle lane markings, particular lighting or bus shelter types. For those cases we train models specifically for your asset list, using multisensor data captured in your own city. The same 11K 360° imagery doubles as training data, and COCO-formatted multisensor datasets are available through marketplace.akys.ai.
See what the models find on your streets
We drive a section of your city and return the real output: AI detections with coordinates, 11K 360° imagery and the point cloud behind them. AKYS works across Lithuania (Vilnius, Kaunas, Klaipėda), the Baltics and the EU. Pricing is quoted per project: write to [email protected] or call +370 677 72373.