Advanced Assistance for Traffic Crash Analysis: An AI-Driven Multi-Agent Approach to Pre-Crash Reconstruction

Technical Deep Dive: Advanced Assistance for Traffic Crash Analysis

Overview

This study presents an AI-driven multi-agent framework to automate the analysis of pre-crash scenarios and identify the first collision event in vehicle crashes. The framework combines two phases:

1. Crash Reconstruction: A multimodal agent processes crash scene descriptions and diagrams to generate a structured reconstruction of the collision sequence.
2. First Crash Inference: A reasoning-specialized agent analyzes the reconstruction and vehicle Event Data Recorder (EDR) data to determine the striking and struck vehicles in the initial collision, and identify the relevant EDR event records.

The framework was evaluated on 277 real-world rear-end lead vehicle deceleration (LVD) crashes from the Crash Investigation Sampling System (CISS) dataset. It achieved 100% accuracy in correctly identifying the vehicles and EDR records for the first collision event, outperforming research analysts without specialized reconstruction training (92.31% accuracy).

Problem & Context

• Traffic collisions remain a major public health and safety challenge, with over 1 million fatalities worldwide each year.
• Traditional pre-crash reconstruction methods rely heavily on manual analysis of fragmented evidence (witness accounts, physical evidence, vehicle data), leading to inconsistencies and reduced accuracy.
• The central challenge is integrating heterogeneous multimodal data (crash narratives, vehicle dynamics, scene diagrams) under conditions of missing or contradictory information to infer the critical pre-crash sequence of events.
• Large language models (LLMs) show promise in addressing this challenge through superior logical reasoning and multimodal data integration capabilities.

Methodology

1. Data Processing:
   • Utilized the Crash Investigation Sampling System (CISS) dataset, which contains crash narratives, vehicle dynamics data from Event Data Recorders (EDRs), and scene diagrams.
   • Transformed raw multimodal data into structured natural language formats to provide explicit contextual relationships.
2. Multi-Agent Framework:
   • Phase I (Crash Reconstruction): A multimodal agent processes the scene description and diagram to generate a structured reconstruction of the collision sequence.
   • Phase II (First Crash Inference): A reasoning-specialized agent analyzes the reconstruction and EDR data to determine the striking/struck vehicles and identify the relevant EDR event record for the initial collision.
3. Prompt Engineering:
   • Designed structured prompts with reasoning anchors to guide the agents through the analytical workflow and maintain consistent judgments across different LLM architectures.
4. Evaluation:
   • Tested the framework on 277 LVD cases, with 238 "simple" (one-to-one event-record mapping) and 39 "complicated" (many-to-one) cases.
   • Compared performance against research analysts without specialized reconstruction training.

Results

• The AI-driven framework achieved 100% accuracy across all 4,155 experimental trials, correctly identifying the striking/struck vehicles and the most relevant EDR records for the initial collision event.
• In the 39 complicated EDR cases, the framework outperformed the research analysts, who achieved 92.31% accuracy.
• A cross-model evaluation showed that three different reasoning agents produced identical outputs, confirming the effectiveness of the structured prompt design.
• Ablation experiments demonstrated that removing the reasoning anchors reduced case-level accuracy from 99.7% to 96.5%, with errors spreading across multiple output dimensions.
• The AI framework processed each case in under 1 minute on average, over 5 times faster than the human analysts.

Interpretation

• The AI-driven framework provides stable, reproducible analytical support, reducing the cognitive load and inconsistencies observed in manual pre-crash reconstruction.
• The structured reasoning anchors were critical in maintaining inference stability, particularly in complex scenarios with ambiguous or contradictory evidence.
• While not replacing expert judgment, the framework can assist non-specialist personnel in systematically screening evidence, reducing omissions, and triaging cases for further investigation.
• The natural language outputs are compatible with emerging generative AI capabilities, potentially enabling automated visual reconstruction of collision processes.

Limitations & Uncertainties

• The evaluation was limited to rear-end LVD scenarios and did not cover other crash types.
• The "ground truth" for the 39 complicated EDR cases was established through a consensus process involving trained researchers, not certified reconstruction experts.
• Future research will focus on extending the framework to additional crash types, incorporating expert-labeled benchmarks, and exploring adaptive reasoning strategies to balance inference stability and computational efficiency.

What Comes Next

• Expand the framework to handle a broader range of crash types beyond rear-end LVD scenarios.
• Integrate expert-labeled benchmarks to further validate the framework's performance.
• Leverage the accurate first collision identification to enable large-scale analysis of pre-crash driver behaviors and vehicle dynamics.
• Explore integrating the natural language reconstruction outputs with generative AI capabilities to produce visual collision process visualizations.
• Develop adaptive reasoning strategies to balance inference stability and computational efficiency for real-world deployment.