TLDR: Researchers developed an AI framework using a fine-tuned Large Language Model (Llama 3.2 1B) to automatically identify Driver Hazardous Actions (DHAs) from textual crash narratives with 80% accuracy, outperforming traditional models. A novel probabilistic reasoning approach simulates hypothetical scenarios (driver distraction, teen drivers) to interpret AI decisions, revealing how these factors shift the likelihood of specific hazardous actions like “General Unsafe Driving” or “Speed and Stopping Violations,” offering new insights for traffic safety analysis and interventions.
Vehicle crashes are complex events, often involving split-second decisions and challenging conditions. Among these, two-vehicle crashes are the most common, making up about 70% of all roadway incidents. Understanding why these crashes happen is crucial for improving traffic safety, and a key part of this is identifying Driver Hazardous Actions (DHAs) – the specific maneuvers or behaviors that lead to a crash.
However, current methods for identifying DHAs in large databases are often inconsistent and require a lot of manual effort. This makes it difficult to get a reliable and clear picture of crash causes. To address this challenge, researchers have developed an innovative framework that uses a powerful artificial intelligence tool called a Large Language Model (LLM) to automatically figure out DHAs from written crash reports.
The study, titled “From Narratives to Probabilistic Reasoning: Predicting and Interpreting Drivers’ Hazardous Actions in Crashes Using Large Language Model”, was conducted by a team of researchers including Boyou Chen, Gerui Xu, Zifei Wang, Huizhong Guo, Ananna Ahmed, Zhaonan Sun, Zhen Hu, Kaihan Zhang, and Shan Bao. They aimed to improve the accuracy and understanding of DHA classifications.
The team used five years of two-vehicle crash data from Michigan Traffic Crash Facts (2019–2023). They took the Llama 3.2 1B model, a type of LLM, and fine-tuned it using detailed textual crash narratives. This means they trained the AI specifically on how crashes are described, along with other structured information like crash details, driver information, road characteristics, and environmental conditions. To make the analysis clearer, the 14 original officer-assigned DHA codes were grouped into seven broader, more understandable categories.
The fine-tuned LLM proved to be highly effective, achieving an overall accuracy of 80%. This performance significantly surpassed traditional machine learning models like Random Forest, XGBoost, CatBoost, and a neural network, especially in situations where some types of crash data were less common.
One of the most exciting aspects of this research is a new approach called probabilistic reasoning. Since LLMs can sometimes be like “black boxes” where it’s hard to see how they make decisions, the researchers developed a way to interpret the model’s outputs. They did this by simulating different hypothetical crash scenarios that are known to be high-risk but might not be common in real-world data.
Also Read:
- Decoding the Road: How Align2Act Brings Human Logic to Self-Driving Cars
- Enhancing Autonomous Vehicle Safety and Efficiency at Unsignalized Intersections with Biased Attention
Understanding Crash Scenarios with AI
The researchers explored three specific counterfactual scenarios:
- Single-Driver Distraction: When one driver was introduced as distracted, the model showed a substantial increase in the likelihood of “General Unsafe Driving.” This suggests that distraction in one driver often leads to broad unsafe behaviors.
- Both-Driver Distraction: When both drivers were labeled as distracted, the model predicted the highest probability for “Both Drivers Took Hazardous Actions.” It also showed an increase in “Lane and Direction Violations,” indicating that dual distraction can lead to more specific directional control issues.
- Teen Driver Scenario: When all drivers were replaced with teenagers (aged 16-17), the model significantly increased the probability of “Speed and Stopping Violations” and “Right-of-Way and Traffic Control Violations.” This aligns with existing knowledge about the elevated crash risks associated with young drivers, particularly concerning speeding and right-of-way issues.
These findings offer a robust and understandable way to detect DHAs automatically on a large scale. This framework opens up new possibilities for analyzing traffic safety and developing interventions. For instance, understanding how distraction or young drivers influence specific hazardous actions can help policymakers design more targeted safety campaigns or improve driver education programs.
In essence, this study moves the field towards richer, more explainable insights into crash causation, which can ultimately lead to more effective safety measures on our roads.
