In road safety management one key issue is allocating otherwise limited resources to address accident hotspots effectively and efficiently. While data about accidents is collected, usually authorities are forced to act after accident hotspots already emerged, instead of preventing accidents.

EvoRoads partner Inlecom (in collaboration with partner CSI-Piemonte for its validation) developed a new AI-Powered Road Accident Risk Prediction and Mapping Tool that addresses exactly this pressing topic.By turning historical accident data that virtually every road authority already collects into actionable safety intelligence, the tool classifies every segment of a road network by current accident risk level, and risk trajectory, highlighting if they are stable, increasing, or decreasing in time. Thanks to such a predictive view, authorities can prioritise and allocate budgets in areas of need, addressing rising risks before they turn into accidents and ensuring ongoing and constant safety for all road users. The tool has been validated in Piedmont, Italy.

How will this tool impact European citizens?

Thanks to this risk prediction and mapping tool, regional road authorities can now generate a complete risk picture of their entire network. In the validation region the tool covers 99.7% of all segments with recorded accidents. Thanks to its design as complementing already carried out road safety activities, the tool does not require any new sensors or infrastructure, and it is expected to measurably reduce crashes, injuries, and fatalities thanks to better-targeted infrastructure interventions, especially in areas with otherwise already limited budgets.

How was the tool developed?

Inlecom built the tool in four steps. 

First, the team created a data pipeline to collect and standardise years of accident data from different sources — crash locations, injuries, fatalities and traffic flow — into one consistent dataset. 

Second, they developed a new way of grouping road segments, dubbed “thickened segments,” which links each stretch of road with its real neighbours, such as junctions and parallel lanes. This ensures the risk picture reflects actual accident hotspots, rather than being distorted by how the road network happens to be mapped digitally. 

Third, the team built and tested the AI models at the core of the tool. These models do three things: group road segments into four risk levels, estimate whether accidents on each segment are rising or falling over time, and classify that trend into ten more detailed categories. Several methods were tested for each task, and only the most reliable ones were kept. 

Finally, the tool was tested at full scale on the road network of Piedmont, Italy, using over 3.5 million accident records from 2016–2022, provided by CSI Piemonte. Local practitioners helped shape the tool’s requirements from the start, and the result was a working demonstrator with interactive maps — designed to help prioritise which road segments and municipalities need the most urgent action.

Want to learn more?

Results are shared openly through multiple channels. The complete methodology for its development is available within EvoRoads public deliverable D2.1 (July 2025), freely available to any organization here. An online demonstrator showcases the tool’s risk maps and trend analysis (link). Furthermore, the tool will be showcased in EvoRoads pilots, events, and other outreach activities.