Traffic signs are among the most ubiquitous and overlooked assets in any city’s infrastructure portfolio. Speed limits, parking restrictions, priority rules, pedestrian zones: thousands of signs govern daily traffic flow, each carrying legal weight and maintenance responsibilities. Yet across Europe, municipal traffic sign databases remain surprisingly fragmented, outdated and incomplete.

The consequences of inaccurate inventories extend beyond administrative inconvenience. When a sign is missing from the database, maintenance schedules skip it. When records show a sign that no longer exists, crews waste time searching for phantom infrastructure. When citizen complaints arrive about unclear or damaged signage, response times stretch while staff manually verify what’s actually installed.

Stad Roeselare recognized an opportunity to break this cycle. Rather than treating traffic sign inventory as a periodic project, they envisioned it as a continuous process, one that could leverage mobile sensing technology already deployed across their fleet or fleet of partners servicing the city.

Traffic signs are among the most ubiquitous and overlooked assets in any city’s infrastructure portfolio. Speed limits, parking restrictions, priority rules, pedestrian zones: thousands of signs govern daily traffic flow, each carrying legal weight and maintenance responsibilities. Yet across Europe, municipal traffic sign databases remain surprisingly fragmented, outdated and incomplete.

The consequences of inaccurate inventories extend beyond administrative inconvenience. When a sign is missing from the database, maintenance schedules skip it. When records show a sign that no longer exists, crews waste time searching for phantom infrastructure. When citizen complaints arrive about unclear or damaged signage, response times stretch while staff manually verify what’s actually installed.

Stad Roeselare recognized an opportunity to break this cycle. Rather than treating traffic sign inventory as a periodic project, they envisioned it as a continuous process, one that could leverage mobile sensing technology already deployed across their fleet or fleet of partners servicing the city.

The transition from concept to deployment revealed challenges familiar to anyone working with computer vision in uncontrolled environments:

• Variable lighting conditions: Morning glare, afternoon shadows, overcast days, … or typical Belgian weather in a nutshell.

• Oblique camera angles: Signs photographed from passing vehicles at unpredictable angles

• Partial occlusions: Parked cars, overgrown vegetation, construction barriers

• Sign degradation: Weathering, vandalism, stickers and fading

• Diverse sign taxonomy: Belgian traffic signs span multiple categories (Warning, Priority, Prohibition, Mandatory, Parking, Additional) with dozens of subtypes

These challenges aren’t unique to Roeselare. As research from the ViCoS Lab [Traffic-sign detection] highlights, automating traffic sign inventory management requires deep learning models that can handle the gap between idealized training data and messy real-world captures.

Working with Stad Roeselare, the Sensr team developed a two-stage AI pipeline: detection on the edge, classification in the cloud.

The team selected YOLOv11n as the base detection model, balancing speed and accuracy for real-time detection during MSU operations. This choice reflects broader industry trends. Research shows YOLO variants consistently outperform other architectures for traffic sign detection, achieving mAP50 scores exceeding 90% in controlled settings [Traffic Sign Detection and Recognition Using YOLO Object Detection Algorithm: A Systematic Review].

Rather than training a single monolithic classifier, the solution employs a hierarchical classification approach with specialized sub-models for categories A (Warning), B (Priority), C (Prohibition), D (Mandatory), E (Parking), and M (Additional).

How it works:

1. The high-level classifier determines the main category (e.g., “This is a prohibition sign”)

2. A specialized sub-model for that category identifies the specific type (e.g., “B1: Priority from the right”)

3. For edge cases like M3bis/M11 signs, a dedicated ensemble model combining optical character recognition (OCR) with pixel analysis provides fine-grained classification

This architecture offers several advantages:

• Specialization: Each sub-model focuses on distinguishing signs within a single category, improving accuracy

• Robustness: If the high-level classifier is uncertain, the system can flag the detection for human review rather than forcing a potentially wrong classification

• Extensibility: New sign types can be added by training additional sub-models without retraining the entire system

Classification accuracy across the hierarchy ranged from 85.7% to 99.7% depending on the category, with the most common sign types achieving the highest performance.

The classification pipeline runs on Microsoft Azure with a serverless architecture designed for cost efficiency:

• FastAPI: Handles inference requests with low-latency response times

• Azure Service Bus: Manages asynchronous message queuing between the API and processing layer

• Inference Processor: Downloads images, applies classification, returns results to Stad Roeselare’s Data Platforms.

• PostgreSQL: Stores structured results with geolocation metadata and observability data via OpenTelemetry

• Scale-to-zero capability: Compute resources scale down when idle, minimizing costs during gaps between MSU deployments

This architecture mirrors patterns we’ve described in our earlier post on A Robust Deployment Pattern for ML Models at the Edge, where we emphasized the importance of designing for unreliable connectivity and variable processing loads.

The deployed system achieved the following detection performance:

The gap between these numbers tells an important story. In controlled conditions with clean images, optimal angles, good lighting, the model performs excellently. In real-world deployment, performance drops but remains practically useful. This pattern is consistent with findings from traffic sign detection research, where models achieving 90%+ mAP in benchmarks typically see reduced performance on in-the-wild data.

For Stad Roeselare, the goal wasn’t perfect detection, it was operational improvement. The system:

• Runs continuously: Every MSU ride automatically updates the inventory

• Reduces manual effort: Staff no longer drive dedicated survey routes

• Enables faster response: Current sign locations are always available for maintenance planning

• Catches changes: New signs, removed signs and relocated signs are detected automatically

As Roadscan’s case studies [RoadScan | Smart Roads Safer Tomorrow] demonstrate, AI-powered traffic sign management can transform month-long manual processes into operations completed in days, with greater accuracy and consistency.

• Continuous inventory updates: Database reflects ongoing MSU captures

• Resource reallocation: Staff focus on higher-value activities like sign maintenance and citizen response

• Reduced survey costs: Eliminating dedicated survey vehicles and crews

• Faster response times: Current imagery and sign locations always available

The decision to use specialized classifiers for each sign category proved crucial. Belgian traffic signs include over 1000 distinct (sub)types. A single model trying to distinguish all of them would struggle with the long-tail distribution. By breaking the problem into manageable sub-problems, each model could focus on the distinctions that matter most.

Training solely on clean databases would have produced a model that failed in deployment. The reference ride footage, despite its lower quality, was essential for teaching the model what signs actually look like from a moving vehicle on Belgian roads.

The most sophisticated model is useless if it can’t feed results into existing workflows. The Sensr team invested heavily in integration with Stad Roeselare’s data platforms, ensuring that classified signs appeared where municipal staff already worked.

This philosophy, technology serving existing processes, not demanding new ones, echoes our work with Electrolux’s EMC² platform, where we emphasized that the human role remains essential in the synergy between technology and human expertise.

The 77% real-world detection rate isn’t a failure, it’s a realistic baseline that the municipality can build on. Overpromising accuracy leads to disappointment, setting honest expectations allowed the project to be evaluated on its actual operational impact.

Stad Roeselare’s traffic sign project is part of a broader shift in how municipalities approach infrastructure management. As research on smart city traffic management shows, AI-driven systems can reduce travel times by 10-20% and significantly cut emissions through optimized traffic flow.

But the technology’s value extends beyond traffic. The same MSU that captures signs could potentially:

• Detect road surface damage for proactive maintenance

• Identify illegal dumping or unauthorized construction

• Monitor parking compliance in restricted zones

• Track vegetation overgrowth affecting visibility

The infrastructure is in place; the limiting factor is often the availability of trained models and integration with municipal workflows.

The Stad Roeselare traffic sign classification project demonstrates that successful AI deployment requires more than model accuracy. It demands thoughtful architecture, seamless integration with existing systems and realistic expectations about edge-case performance.

Most importantly, it shows that practical AI beats perfect AI. A system that runs continuously at 77% accuracy delivers more value than a hypothetical 99% system that never deploys. For municipalities looking to modernize infrastructure management, the lesson is clear: start with what you have, build what works, and improve iteratively.