Learning Pedestrian Failure-to-yield Maneuver Patterns from Fatal Crash Data: Evidence from Explainable AutoML

Every day, millions of people step off the curb to cross a street, trusting that drivers will see them and stop. Yet, in cities across the United States, this trust is frequently broken. When a pedestrian crosses without waiting for a safe gap or ignoring a signal, and a vehicle fails to yield, the result is often a tragedy. These "failure-to-yield" incidents are not random accidents; they follow specific patterns based on where the person is standing, what time of day it is, and who is behind the wheel. For decades, safety experts have tried to understand these moments by manually reviewing police reports or watching video footage, but these methods are slow and struggle to handle the sheer volume of data available today. The question remains: can we use the massive archives of traffic records to uncover the hidden rules that govern these dangerous interactions, and can we do it in a way that explains why a crash happened, not just that it did?

A team of researchers at Texas State University set out to answer this question by turning to a new kind of computer intelligence. They gathered data from the Fatality Analysis Reporting System, a national database containing detailed records of every fatal traffic crash in the United States between 2016 and 2023. Instead of trying to predict a single number, like the total number of deaths, they asked the computer to sort these tragic events into specific categories based on how the pedestrian was moving at the moment of impact. They used a classification system known as the Pedestrian and Bicycle Crash Analysis Tool, which breaks down crashes into types such as a person crossing from the left, crossing from the right, crossing in an unusual direction, or crossing with an unknown direction. By feeding this data into an automated machine learning system called AutoGluon, the researchers allowed the computer to build its own models to find the most accurate way to distinguish between these different types of crashes.

The computer did not just guess; it learned. After testing dozens of different mathematical approaches, the system settled on a method called LightGBM, which proved to be the most accurate. This model correctly identified the type of pedestrian maneuver in nearly 80 percent of the fatal cases it was tested on. More importantly, the researchers did not leave the computer as a "black box" that simply gives an answer without explanation. They used a technique called SHAP, which acts like a magnifying glass, to look inside the model and see which factors were pushing the prediction one way or the other. This revealed that the year the crash occurred, the exact position of the pedestrian, the type of road, and the age of the driver were the most powerful clues in determining how a crash unfolded.

When the researchers looked closely at the results, distinct stories emerged for each type of crossing. For crashes where a pedestrian crossed from the left, the computer showed that the most critical factor was the pedestrian's position. These incidents were heavily linked to people crossing in the middle of a block, away from marked crosswalks, where drivers do not expect them to appear. The analysis also highlighted that older drivers were particularly vulnerable in these scenarios, likely because their reaction times slow down when faced with a sudden, unexpected person stepping out from the side.

The story was different for crashes involving a pedestrian crossing from the right. Here, the age of the driver again played a major role, but so did the type of vehicle and the time of day. The data suggested that these crashes often happen when a driver is making a right turn, a maneuver that requires looking in multiple directions at once. The computer's analysis indicated that drivers often fail to scan the area where a pedestrian might step out, especially if the view is blocked by the car's own pillars or if the road design includes slip lanes that allow cars to turn without fully stopping. In these cases, the driver's attention is divided, and the pedestrian is hidden in a blind spot until it is too late.

For the most confusing category, where the direction of the pedestrian was unknown or unusual, the pattern pointed toward older pedestrians and poor visibility. These crashes frequently involved people crossing in the middle of the night or in bad weather, often far from any designated crossing point. The data showed that these incidents were more common in the winter months and often involved larger vehicles like SUVs, which can block a driver's view of a person on the ground. The computer learned that when a pedestrian crosses in an unpredictable way, the combination of darkness, weather, and the size of the vehicle creates a perfect storm for a fatal error.

The findings challenge the idea that all pedestrian crashes are the same. Instead, they show that each type of crossing failure has its own unique set of causes. A crash caused by a jaywalker in the middle of a block is driven by different factors than a crash caused by a pedestrian stepping out during a right turn. This distinction is vital for safety planners. It suggests that a single solution, like painting more crosswalks, will not fix every problem. To truly save lives, cities may need to redesign specific intersections to improve sight lines for turning drivers, install better lighting for night crossings, and create targeted education programs for older drivers who may struggle to react quickly to unexpected pedestrians.

By using automated learning to sift through years of fatal crash data, this study has moved beyond simple statistics to reveal the specific mechanics of danger. It proves that with the right tools, we can understand the precise conditions that lead to tragedy. The research does not offer a magic fix, but it provides a clear map of where the risks lie. It shows that safety is not just about following rules, but about understanding the complex dance of human behavior, vehicle design, and road geometry that happens in the split second before a crash. With this knowledge, engineers and policymakers can finally begin to build streets that are not just safer for the average day, but resilient enough to protect the most vulnerable when things go wrong.