From Waze alerts to actionable intelligence: detecting road incidents 48% faster

A national highway operator manages incident detection and road assistance across highways through its Control Centre.

Incidents are detected through multiple sources: road assistance vans account for the majority of incidents, with the remaining coming from CCTV cameras, and user reports through the call centre. There is a high coverage of CCTV in the network, however it varies by location and highway.

The challenge: detecting road incidents faster without increasing false positives

Incident detection remained heavily dependent on road assistance vans, while crowdsourced Waze alerts generated too many false positives to be used reliably without manual CCTV validation.

The objective was therefore to answer a practical question: Can machine learning make crowdsourced Waze data reliable enough to detect more road incidents, detect them earlier, and keep false positives at an operationally manageable level?

The solution: machine learning for real-time road incident detection

LTPlabs developed a machine learning solution that evaluates crowdsourced Waze events and estimates which alerts are sufficiently reliable to support road incident detection and monitoring.

The approach combines Waze data with contextual information about traffic, weather, road attributes, location, time, and historical incidents.

The first step was to match Waze alerts with incidents registered by the Control Centre. Events were grouped based on location, time, and direction of travel, allowing 79% of registered incidents to be matched with Waze data.

Using this historical data, LTPlabs trained a classification model to estimate the likelihood that a Waze alert corresponded to a relevant incident. The model combines 20+ variables, including Waze attributes, traffic and weather conditions, road characteristics, location, time, and accident history.

Confidence thresholds were then calibrated to balance earlier detection with the risk of false positives, with different thresholds proposed depending on CCTV coverage.

When an alert reaches the required confidence level, it can be automatically sent to Control Centre supervisors and integrated into the existing dispatch workflow, enabling faster validation and systematic performance tracking.

Results: towards 48% faster road incident detection

On the holdout test set, and following a pilot in a real operational context, the selected operating thresholds achieved 83% accuracy and 61% more alerts compared with baseline assumptions. The solution then creates a path to reducing average incident detection time by 48% of a minute, and combined with existing detection sources, Waze alerts are also expected to increase registered events significantly.

The analysis demonstrates that crowdsourced traffic data can become a useful road incident detection source when machine learning is used to filter alerts and combine them with contextual information.

For highway operators, this creates the potential to:

• detect road incidents earlier;
• extend monitoring beyond areas with CCTV coverage;
• reduce dependence on patrol-based detection;
• prioritise supervisor attention toward higher-confidence alerts;
• improve consistency in incident detection and performance measurement;
• expand network visibility without a proportional increase in physical monitoring infrastructure.