University Professor Uses Artificial Intelligence to Improve Road Safety

A University of Houston civil engineering professor is using artificial intelligence to help agencies make pavement-related decisions that can reduce crash risk. The work links roadway condition records—normally analyzed separately from crash reporting—to large-scale police crash narratives, using large language models to reveal details often missing from structured fields.

In a Texas Department of Transportation–funded study, Lu Gao integrated multiple data streams describing pavement structure, surface condition, and roadway geometry with crash records drawn from police reports. The centerpiece of the approach is narrative text analysis: officers’ descriptions of events were processed to extract crash “mechanisms” such as hydroplaning or loss of control on curves.

Traditional crash databases capture many outcomes, but they frequently lack pavement- and surface-specific context. That gap matters because pavement texture, friction, and surface condition influence both how crashes occur and how severe they become, particularly under wet conditions. By translating unstructured text into consistent labels, the AI makes those pavement-linked mechanisms measurable at scale.

Using more than 24,000 police crash narratives aligned to roughly 180,000 pavement-management records, the study reports strong associations between friction/texture indicators and wet-pavement crash mechanisms. In practical terms, the results point to which roadway segments may be most responsive to pavement-focused interventions rather than broad, non-targeted safety spending.

The methodology combines probabilistic interpretation with mechanism-specific categorization, enabling quantification of how specific pavement properties correspond to distinct crash pathways. This supports a screening logic that goes beyond “where crashes happen” and toward “why crashes happen,” grounded in surface condition evidence.

Because the extraction relies on language understanding instead of rigid keyword rules, it can capture varied narrative phrasing across thousands of incidents. That makes the workflow adaptable to real agency datasets and potentially transferable to other regions with narrative-based reporting.

Ultimately, the research aims to give transportation departments a more evidence-driven way to prioritize maintenance and safety treatments. When pavement and crash mechanisms are connected, agencies can target cost-effective countermeasures to segments where elevated crash risk is more strongly linked to pavement conditions.

The findings are published in Accident Analysis and Prevention, demonstrating that LLM-assisted narrative parsing can enhance pavement-safety assessment without waiting for costly manual review at highway scale.

Keywords

Applied sciences and engineering; civil engineering; pavement safety; artificial intelligence; large language models; crash narrative analysis; transportation safety; pavement friction/texture; wet-pavement mechanisms