RiskGrid Turns Simulated Traffic Into Road Safety Risk Maps

The study, published in European Transport Research Review, addresses a long-standing blind spot in urban planning. Mobility Impact Assessments, the structured analyses required when new developments are proposed, traditionally focus on operational and environmental indicators such as volume-to-capacity ratios and travel times. Road safety, the authors argue, remains peripheral and is seldom addressed during early design stages. That omission is particularly hazardous in expanding suburban areas, where new residential settlements push heavier traffic onto corridors with complex intersections and high speeds, and interventions that look harmless for traffic flow can quietly worsen network safety. European Directive 2019/1936 has pushed member states toward proactive Road Safety Impact Assessments, and RiskGrid is designed to fill that methodological gap.

The framework was tested on the Morena neighborhood, a rapidly growing suburban district on Rome’s southeastern periphery, bounded by major arterial corridors including Via Tuscolana and Via Anagnina. The planned residential development there comprises a gross floor area of 20,160 square meters and would accommodate roughly 538 residents. To keep the assessment conservative, the researchers adopted a worst-case assumption, assigning 100 percent of the estimated 80 peak-hour passenger-car-equivalent trips generated by the development to private motorized transport. Baseline traffic data were collected with a non-intrusive K-Band 24 GHz Sierzega SR4 radar during weekday morning peak hours in October, capturing vehicle-by-vehicle speed, length and time-gap measurements across 11 network sections for model calibration.

The methodology runs in three stages. First, a macrosimulation of supply and demand was built in TransCAD, with the origin-destination matrix iteratively adjusted using the Nielsen algorithm, a Bureau of Public Roads volume-delay function and user-equilibrium assignment, achieving a statistical fit of R² = 0.98. The macroscopic output was then converted into the SUMO microsimulation environment through automated Python scripts. Notably, the team used a Breadth-First Search algorithm rather than iterative dynamic assignment to generate microscopic routes, sacrificing some behavioral realism to keep the pipeline fully autonomous and fast. Longitudinal vehicle dynamics followed the Intelligent Driver Model, with SUMO’s standard LC2013 lane-changing logic handling lateral movements. The microsimulation replicated observed queuing with R² = 0.78 and exported sub-second floating car data trajectories. The authors caution that errors in the macroscopic demand estimate can cascade through route generation and compound in the simulated vehicle behaviors.

RiskGrid’s core idea is a conceptual shift from counting discrete conflicts to mapping continuous risk. The study area is discretized into square cells — 10 by 10 meters in the case study — and vehicles are assigned to cells by their projected coordinates. The tool’s premise is that conflict risk is fundamentally tied to directional heterogeneity: when flows intersect, crossing interactions, merges and turning maneuvers become far more likely. To quantify this, approach angles are grouped into 30-degree bins and combined into a Multi-Directionality Index, a convex blend with equal weighting of normalized entropy, which captures how spread the flow is across angular sectors, and circular variance, which measures the geometric dispersion of trajectories. The index is then scaled by a severity-opportunity factor that weights collision types by expected kinetic severity — head-on conflicts weighted 0.45, angle conflicts 0.30 and rear-end conflicts 0.25 — producing a composite risk index between zero and one.

The output is interoperable with geographic information systems: risk values are assigned Well-Known Text geometries and segmented into five priority classes using a right-skewed quantile distribution at the 50th, 75th, 90th and 97th percentiles, visually isolating the most critical hotspots. Preliminary exploratory testing across a lambda range of 0.4 to 0.6 showed consistent spatial prioritization of high-conflict cells, suggesting hotspot rankings are robust to moderate parameter shifts. The severity weights derive from standard kinetic severity hierarchies in the road safety literature rather than site-specific regression, which the authors say supports generalizability to data-scarce environments, though full empirical sensitivity analyses are planned for future iterations.

Validation against historical crash records from Italy’s national statistics institute, covering 2021 to 2023 with pedestrian and micromobility crashes excluded, exposed a stark contrast. SSAM3, the Federal Highway Administration tool that computes Time-to-Collision and Post-Encroachment Time indicators from trajectories, correctly identified only two true hotspots while generating three false positives and missing 11 critical locations. It produced elongated, diffuse conflict lines that failed to match the localized nature of crash black spots, and it systematically undervalued crossing conflicts — paradoxically the most frequent and severe crash type in the peri-urban study area. RiskGrid, by contrast, detected 12 of 13 hotspots with five false positives and a single false negative. The authors argue the extra false alarms are strategically acceptable under the principle of cost asymmetry: missing a lethal hotspot is effectively infinitely costlier than inspecting an intersection that proves safe.

Computational efficiency told a similar story. Because SSAM3 analyzes every individual vehicle-to-vehicle interaction and requires manual trajectory conversion and post-processing, it took over 12 minutes per scenario. RiskGrid’s automated Python pipeline processes spatially aggregated values instead, generating final classified risk maps for an entire scenario in less than one minute — a saving the authors say becomes highly relevant for multi-scenario comparisons and regional-scale planning.

Applied to compare the do-nothing and project scenarios, RiskGrid revealed what traditional mobility metrics had missed. The development’s 80 additional peak-hour trips were fully absorbed by the network’s residual capacity: average speeds fell a negligible 0.22 percent, from 44.7 to 44.6 km/h, vehicle-kilometers and vehicle-hours rose only marginally, and no new bottlenecks emerged. By conventional assessment standards the project was operationally neutral. Yet RiskGrid calculated a 2.1 percent increase in overall network risk. Assuming the simulated risk scale translates proportionally to observed crashes, that corresponds to roughly two additional expected collisions in the evaluation area, raising the total from 84 to 86 — a finding the authors present as evidence that congestion-only assessments can conceal measurable safety degradation.

The researchers are candid about limitations. The current model evaluates only motorized vehicles, excluding pedestrians and cyclists; it uses a two-dimensional network representation, so trajectories on grade-separated structures such as underpasses are erroneously projected onto the same plane and perceived as intersecting; validation relied on just three years of crash data and only 13 hotspot locations, which limits the statistical power of precision and recall estimates; the tool was applied only to the Morena area, so generalizability to cities with different infrastructure or traffic dynamics remains untested; and the analysis requires a specific floating car data structure, constraining interoperability. Future work will add pedestrian and cyclist trajectories, incorporate a vertical dimension to resolve grade-separated junctions, and extend testing to contexts from dense historic centers to rural highways. Even so, the study makes a concrete case that continuous, exposure-based risk mapping can be folded into routine planning workflows — giving municipalities a rapid pre-screening tool to flag high-risk nodes, prioritize road safety inspections and demand design changes before projects are approved, in line with the Safe System Approach.

Subject of Research: A data-driven framework that maps road crash risk from simulated traffic trajectories to support proactive road safety impact assessment

Article Title: Development of a risk map analysis tool (RiskGrid) to support road safety impact assessment

Article References: Michiorri, V., Gonzalez-Hernandez, B., Usami, D. S., & Persia, L. (2026). Development of a risk map analysis tool (RiskGrid) to support road safety impact assessment. European Transport Research Review, 18(1), Article 100. https://doi.org/10.1186/s12544-026-00864-8

Image Credits: AI Generated

DOI: 10.1186/s12544-026-00864-8

Keywords: road safety, RiskGrid, traffic simulation, microsimulation, SSAM3, Mobility Impact Assessment, SUMO, crash hotspots, surrogate safety measures, urban planning, Rome, Safe System Approach