Intelligent Road Infrastructure

​

​

Sofia, October 2025

wisertech.com

Project Overview

AI-based analysis of very high resolution optical Space imagery with focus on road feature extraction.

The project lays the foundation for building a Digital Twin of the road infrastructure in Bulgaria.

​

​

Intelligent Road Infrastructure - Goals & Impact

Overview

• Empowering Data-Driven Decisions
• Road Quality Assessment
• Automated Infrastructure Risk Analysis

​

​

​

Goals & Impact

Leveraging Satellite Imagery (Airbus Pleiades)

​

Leveraging Satellite Imagery (Airbus Pleiades)

Benefits

• High resolution (50 centimeters per pixel)
• Global coverage
• Commercial license

​

Limitation

• High Cost
• Complex Image Acquisition
• Resolution restricts us from seeing finer details

AI for Road Analysis

• Road Detection and Segmentation
• Bridge Detection and Segmentation
• Road feature extraction
• Road Markings Classification
• Lane Count Classification
• Road Width
• Turns
• Intersections

​

​

​

Key Project Phases

Intersection Detection

Turn Detection

​

Road Width Extraction

Dataset creation for classification

Road marking visibility and lane count classification

​

Dataset creation for Segmentation

Road Segmentation models

Road Detection

Defining "Road"

• Impervious surface that serves as a standard place on which cars must be driven

​

• Roads hidden behind trees and building
• Dirt roads - hard to spot, easy to misclassify, not enough value

​

​

What is not a road?

Dataset creation

• Preprocessing
• Pansharpening
• Annotation - CVAT
• Multi-stage approach
• Splitting

Semantic Segmentation

AI technique that assigns a distinct label to every single pixel in an image.

Road Segmentation classes:

• Road (pink) - all impervious surface a car can be
• Bridge (purple) - all roads that are also bridges
• Parking (yellow) - all impervious parking lots, parking places on the side of the road
• Unclear - if we know there's a road there but it’s not visible

Results from models:

Road Feature Extraction

Feature extraction approach

Split each road into segments with length 50m. Enrich each segment with specific features by:

1. Building datasets out of the segments, manually annotating them & training classification models
2. Building analytical methods for analysis of geometric road properties

Classification with two tasks

Lane Count

Classes:

• Two - road has one or two lanes
• More - road has more than two lanes

Road Marking Visibility

Classes:

• Yes - markings are clearly visible
• No - markings are faint, missing, or unclear

Marking visibility: yes, Lane Count: two

Marking visibility: yes, Lane Count: more

Marking visibility: no, Lane Count: two

Marking visibility: no, Lane Count: more

Road Width Extraction

Use the centerline of the road segment

To calculate the width of each 50m segment:

• Choose a measurement point on each segment
• Draw perpendicular lines at the measurement points
• Measure the intersection length

Turn Detection

• A centerline is defined by many points
• For every 3 consecutive points, we calculate the circumradius of the triangle they form.
• Only points with tight curves (below a set turn threshold) are marked as turn points.
• Nearby points are grouped into Turn objects with an average radius

Intersection Detection

Two Complementary Detection Methods:

1. Endpoint Proximity Analysis:

• Collect all centerline endpoints
• Group nearby endpoints
• Groups with 3+ points → Intersection Point

​

​

2. Line Intersection Analysis

• Compare all pairs of centerlines
• Find geometric intersections (points or overlaps)
• Convert intersection areas to single point locations

​

​

Results from Road Feature Extraction

Discussions

Thank you for your attention!

New York

Chrysler Building,

405 Lexington Avenue,

New York City,

10174 NY�1-727-424-9938

​

Frankfurt

Barckhausstraße 1

60325 Frankfurt am Main

Germany

+49 151 148 65 106

Plovdiv

6, Beethoven str.,

4000 Plovdiv, Bulgaria

Nis

Nikole Pašića 28, IV floor,

18000 Niš, Serbia