Smart Program Card: AI-powered traffic crash prediction and police resource optimisation for Brazil's federal highways

The problem

Brazil's federal highway network spans the entire national territory, and traffic crashes on these roads are a persistent and serious public safety challenge. The frequency of crashes varies considerably depending on factors such as traffic volume, road infrastructure, weather conditions, and regional characteristics.

In the past, decisions about where to deploy police patrols were informed by traditional analytics tools and operational managers' experience. Each decision was made manually based on historical data and the operational knowledge of local commanders. While this approach drew on real expertise, it was limited in its ability to anticipate where incidents were most likely to occur on any given day or to systematically optimise the positioning of patrol resources across large and varied road networks.

The solution

The Brazilian Federal Highway Police, working with researchers at the University of Pernambuco (Brazil) and the University of Exeter (United Kingdom), developed an AI-powered tool that predicts where traffic crashes are most likely to occur and recommends how to position police patrols in response.

The system works in three stages.

First, it uses network science to represent the road system as a graph, in which roads and intersections are connected. This allows the system to capture the relationships between different parts of the highway network rather than treating each road segment in isolation.

Second, it applies graph neural networks, a type of AI specifically designed to learn from connected data. Instead of analysing each road independently, the model considers how conditions on neighbouring roads influence one another to predict which segments are at the highest risk of a traffic crash on a given day.

Third, the system uses optimisation algorithms, known as metaheuristics, to determine the best patrol routes and resource allocation based on those predictions. These algorithms balance multiple objectives at once, such as maximising coverage of high-risk areas while minimising travel time and operational costs.

In simple terms, the AI predicts where incidents are more likely to happen, and the optimisation determines the best way to position police resources to respond proactively.

The current version uses historical traffic crash data, time-related information and road characteristics to generate its predictions. Future versions are planned to incorporate additional data sources such as weather conditions, traffic volume, and other contextual information.

The primary users are operational police commanders responsible for planning patrol activities and allocating resources within their jurisdictions. The tool supports their decision-making by providing data-driven recommendations, while final decisions remain under human responsibility.

The project began as a proof of concept developed in collaboration with a Microsoft partner, alongside PhD research already at an advanced stage. This academic foundation helped accelerate the tool's development and ensured it was built on an extensive review of previous research and several earlier studies conducted by the research group.

The impact

1. Pilot deployed in three regions of Brazil
The tool was deployed in a pilot phase between May and June 2026 across three different regions. Based on the system's predictions, patrol teams were proactively positioned in locations identified as having the highest risk of traffic crashes. Recommendations were prioritised according to the model's confidence level, allowing police resources to be deployed more strategically than under the previous manual approach.

2. Early indicators show reductions in crashes, injuries, and fatalities
The main indicators being monitored are reductions in traffic crashes, injuries, and fatalities in the areas where the tool has been applied. Early results from the pilot have been encouraging. The team has also observed additional operational benefits such as faster response times and improved traffic flow after incidents, although these outcomes have not yet been formally measured.

3. Expansion to ten regions planned
Based on the positive results from the pilot, the Federal Highway Police is planning a second phase to expand deployment to 10 regions over the next months. The long-term goal is to deploy the solution across the entire federal highway network, adapting it to the specific characteristics of each region.

Lessons learned

• Translating a real-world problem into a computational model was the hardest part. Traffic crashes are influenced by many interconnected factors, making them inherently difficult to predict. The most challenging aspect of the project was capturing these underlying relationships and patterns in a form that the AI tool could work with.
• Road segment size matters more than expected. Choosing the appropriate level of segmentation, how long each road section should be for the purposes of prediction, has a significant impact on both the accuracy of predictions and the quality of optimisation recommendations. The team would revisit this decision if starting again, though the extensive earlier research on which the project was built helped avoid many common pitfalls.
• Officer confidence grew as the system proved reliable. Initially, there was some uncertainty among officers about relying on an AI-based decision support tool. However, as the system consistently produced reliable recommendations, confidence grew. Officers now see the tool as valuable support for planning their operations, while final decisions remain under human responsibility.
• The framework performs best where it is needed most. Research findings showed that the framework yields greater improvements in areas with more dispersed road connectivity, suggesting that automated spatial analysis can provide particularly valuable insights into resource positioning in decentralised or less well-connected parts of the network.
• Using only publicly available data was a deliberate design choice. One of the main design decisions was to use only publicly available traffic crash data that does not contain any personal information. This ensured compliance with data protection requirements, allowed the research to be published openly, and facilitated collaboration with universities and other research institutions.

The project was designed to use only publicly available traffic crash data containing no personal information. This ensured compliance with data protection requirements and allowed the research to be shared and reproduced openly without privacy concerns, supporting collaboration with universities and other institutions.