Artificial intelligence can help identify dangerous roads, but even sophisticated prediction models sometimes make the same mistakes. New research published in Accident Analysis & Prevention examines why these errors occur and how identifying their underlying causes could improve pedestrian safety.
By combining computer analysis with human observation, researchers have developed a framework that pinpoints locations where conventional models overlook important road conditions.
<h3>Finding Patterns In Prediction Errors</h3>
Researchers from New York University and the University of Washington analysed 13,706 intersections in Seattle to investigate systematic errors in pedestrian crash prediction.
Rather than evaluating models solely by their overall accuracy, the team examined where their predictions repeatedly went wrong. Their framework combined three indicators: spatial clustering, which identifies geographically concentrated errors; temporal persistence, which reveals mistakes occurring repeatedly over time; and ensemble consensus, which measures agreement between different prediction models.
Together, these indicators helped distinguish potentially correctable modelling problems from less predictable variations in crash data.
<h3>A Small Number Of Troubled Intersections</h3>
The analysis identified 141 high-priority intersections, representing just 1.03% of the locations examined. Remarkably, these sites accounted for 6.6% of the total prediction error.
The researchers developed a combined priority score to identify locations where errors appeared particularly systematic.
The flagged intersections had approximately twice the baseline crash rate, yet the models consistently overestimated their risk. This suggested that the algorithms recognised genuinely hazardous locations but lacked sufficient information to determine their precise level of danger.
The findings highlight an important distinction: identifying a potentially dangerous intersection is not the same as accurately measuring its risk.
<h3>What The Algorithms Overlooked</h3>
To investigate the recurring errors, the researchers introduced human observation into their analytical process. Using Google Street View, they examined high-priority intersections to identify environmental characteristics missing from the original prediction models.
Their investigation revealed three categories of overlooked information: visual obstructions, static road use and network complexity.
These characteristics can influence how pedestrians and drivers interact with their surroundings. Obstacles may restrict visibility, while complicated road layouts and surrounding land use can create conditions that are difficult to represent through conventional datasets.
The researchers found that different models often made similar mistakes at the same locations, suggesting that the problem involved missing information rather than weaknesses unique to a particular algorithm.
<h3>Combining Human Insight With AI</h3>
After identifying the overlooked environmental characteristics, the team incorporated additional information into its predictive framework.
The resulting validation showed that priority scores declined by 6.1% at high-priority intersections, indicating an improvement in the previously identified problem areas.
The researchers suggest that combining automated error detection with targeted human observation offers a practical way to improve road safety models.
Instead of manually inspecting thousands of intersections, transportation specialists could concentrate their attention on locations where several models repeatedly produce inaccurate predictions.
<h3>Improving Future Road Safety</h3>
The findings offer a potential approach to making pedestrian safety assessments more informative. However, the analysis focused on Seattle, and additional research would be needed to establish how effectively the framework performs in other cities.
The 6.1% improvement also represents a reduction in the researchers' priority scores, not a measured decrease in pedestrian collisions.
Nevertheless, the study demonstrates how investigating prediction errors can reveal missing information that ordinary accuracy measurements might overlook. Rather than treating every incorrect prediction as an unavoidable limitation, researchers can use recurring mistakes to identify weaknesses in their data and improve how future safety risks are assessed.
