AUTONEWS

Why pedestrian deaths keep rising: AI spots rare crash patterns where targeted fixes could save lives

On average, car crashes cause more than 40,000 deaths per year in the United States. Technologies like seat belts, advanced airbags, and automated braking systems have improved car driver and passenger safety, but pedestrian deaths due to crashes have actually increased by 48% over the last decade, reaching about 7,500 fatalities in 2022. Transportation researchers comb through police crash reports to identify infrastructure countermeasures that will help in the greatest number of cases. However, sometimes improving the average situation isn't enough.

"By using traditional analysis methods that focus on the average, most studies on pedestrian safety overlook crashes that are rarer but may cause disproportionately high-risk injuries," said Zeinab Bayati, a Ph.D. student in the Department of Civil and Environmental Engineering (CEE) at the University of Tennessee, Knoxville.

The most common pedestrian crashes occur during the day and at intersections, but such crashes also usually result in comparatively minor injuries. High-risk injury scenarios—those in which pedestrians are much more likely to be seriously injured or killed—are less common but still vital to consider, said Bayati's doctoral advisor, CEE Beaman Professor Asad Khattak.

"Let's say it's nighttime, there's torrential rain, and maybe the pedestrian and the person who is driving are both under the influence of alcohol," said Khattak. "Those kinds of situations result in more dangerous crashes, but they are so far outside the average that researchers might even remove them from the data set as outliers to make the general trends clearer."

Using AI methods, Bayati and Khattak have developed a novel framework that analyzes pedestrian crash data and sorts events into meaningful groups. Their research, published in the journal Accident Analysis and Prevention, reveals that safety measures aimed at the most common crashes might not save the most lives.

"We identified that the rare cases are indeed the more fatal cases," Bayati said. "That is very important. We want to see what's going on behind this pattern."

AI reveals commonalities in outliers...Bayati and Khattak used an unsupervised clustering algorithm to analyze the factors involved in more than 10,000 police-reported pedestrian crashes. Each report includes detailed information like the speed limit, lighting conditions, road surface conditions, and pedestrian position at the time of the crash.

They then directed the AI to divide the crashes into three categories relating to their distance from the core, or the cluster of "average" scenarios. While only 8% of core cases were fatal, nearly 37% of cases in the furthest edge category involved fatalities.

Analyzing the factors involved in those life-threatening "outlier" crashes reveals the types of interventions that may be most effective at saving pedestrian lives.

"You can check the history of different locations and look for trends that lead to most crashes there," Bayati explained. "We have had a lot of crashes that happened on road shoulders, so we can consider installing sidewalks, speed bumps, or crosswalks in those locations to give pedestrians the opportunity to walk in safety."

Strategic interventions save lives...Part of what makes the edge cases so dangerous—and so rare—is that they occur when many risky factors coincide. That also means it is harder to design interventions against them.

For example, installing a sidewalk along an unlit stretch of a rural roadway will only make it safer during the day. Darkness was a common factor in the highest-risk crashes Bayati and Khattak analyzed, indicating that installing lighting with a sidewalk would lead to an even greater increase in safety.

These complex and rare scenarios can also be used to improve autonomous vehicle safety by exposing failure modes unlikely to appear in routine driving data. Autonomous cars need to be able to reliably sense pedestrians on the road shoulder even in darkness, for example.

"Nighttime crashes or freeway crashes require a lot of different kinds of countermeasures that are harder to implement," said Khattak. "That might be another reason why safety improvements tend to focus on the most prevalent problems, like at urban intersections."

Bayati hopes that this study will create a better understanding of how pedestrian risk is distributed across a transportation system and result in more impactful safety interventions.

"Pedestrian safety is a very important topic. It affects everyone," she said. "I would be happy to have even a small role in creating a safer transportation system."

AI identifies rare flaws by detecting subtle deviations in large volumes of data that would go unnoticed by human experts. In critical areas such as healthcare and engineering, this technology allows for targeted corrections before problems become fatal. Methods for Identifying Rare Faults To find these "invisible" patterns, AI uses specific approaches:

Anomaly detection (Unsupervised Learning): Algorithms such as Isolation Forest isolate data points that deviate significantly from the norm without needing previous examples of failure. This is vital for detecting new or ultra-rare defects.

Transfer learning: Because data on rare flaws is scarce, models are first trained on large, common datasets and then fine-tuned to recognize the specific signals of rare conditions.

Increased synthetic data: Generative models create "fictitious but realistic" data to expand small datasets, allowing AI to learn failure patterns that rarely occur in real life.

Challenges and the "human factor"...Despite its accuracy, AI in high-risk environments requires human oversight. The opacity of some models ("black boxes") can lead to mistrust or error if the doctor or engineer does not understand the logic behind an alert. Therefore, the use of explainable AI (XAI) is essential for experts to validate machine suggestions before acting.

Provided by University of Tennessee at Knoxville