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

This study presents a zero-shot, multi-agent AI framework that successfully reconstructs pre-crash scenarios and infers vehicle behaviors from fragmented multimodal data, achieving 100% accuracy on 4,155 trials and outperforming non-expert analysts in complex rear-end collision cases without requiring domain-specific training.

The Gist

Imagine you are a detective trying to solve a car crash. Usually, you have a clear photo of the crash site, a witness statement, and a black box (like an airplane's) that records the last few seconds of the car's life. But what if the photo is blurry, the witness is confused, and the black box has recorded three different "crashes" in the last 10 seconds, only one of which is the real one?

That is the nightmare of pre-crash reconstruction. It's trying to figure out exactly what happened before the bump, often with messy, broken, or conflicting clues.

This paper introduces a new AI Detective Team designed to solve these messy puzzles faster and more accurately than a single human could.

Here is how it works, broken down into simple concepts and analogies:

1. The Problem: The "Jigsaw Puzzle" with Missing Pieces

Traditionally, human experts try to rebuild a crash by looking at:

• Text: Police reports and witness stories (which can be wrong or vague).
• Images: Diagrams of where the cars were.
• Data: The "Black Box" (EDR) from the car, which records speed and brakes.

The problem is that these clues often don't match. The witness says "I hit the brakes," but the black box says "I was accelerating." Or, the black box has recorded five different events, and the human has to guess which one was the actual crash. This is hard, tiring work, and humans make mistakes when they are tired or confused.

2. The Solution: A Two-Person Detective Squad (Multi-Agent AI)

Instead of hiring one super-smart AI to do everything, the researchers built a team of two specialized AI agents. Think of it like a law firm where one lawyer is great at reading the scene, and another is great at logic puzzles.

Phase 1: The "Scene Setter" (The Visual Detective)

• Role: This agent looks at the messy diagrams and reads the police reports.
• Superpower: It's like a forensic artist. It takes the confusing sketch of the crash and the messy text report and draws a clear, step-by-step story of what happened.
• Goal: It ignores the noise (like logos on the diagram) and creates a clean, written narrative: "Car A was slowing down, Car B was speeding up, and they were 50 feet apart."

Phase 2: The "Logic Master" (The Time-Traveler)

• Role: This agent takes the clean story from Phase 1 and compares it to the car's "Black Box" data.
• Superpower: It's like a time-traveling judge. It looks at the timeline of data and asks: "The story says the crash happened at 2:00 PM. The Black Box has three events at 2:00 PM. Which one is the real one?"
• The Secret Weapon: This agent follows a strict set of "Reasoning Anchors." Imagine these are like rules of the road for the AI.
  • Rule 1: "If the data is missing, don't guess; say 'No Record'."
  • Rule 2: "If the Black Box says 'T=0' is the crash, remember that might be a split second off."
  • Rule 3: "If two cars have data, cross-check them to see who hit whom."
  • These rules stop the AI from "hallucinating" (making things up) and force it to think logically.

3. The Results: The AI vs. The Humans

The researchers tested this team on 277 real car crashes.

• The Simple Cases: When the clues were clear, both the AI and human experts got 100% right.
• The Hard Cases: They picked 39 really messy crashes where the data was confusing (like the Black Box having multiple conflicting records).
  • Human Experts (without special training): Got it right 92% of the time. They got confused by the messy data.
  • The AI Team: Got it right 100% of the time. Every single time.

The Speed Factor:

• The AI team solved a complex crash in less than 1 minute.
• A human expert took about 6.5 minutes.
• The AI is 7 times faster and never gets tired or frustrated.

4. Why Did the AI Win? (The "Reasoning Anchors")

The researchers did a cool experiment. They took the AI team and told the "Logic Master" to ignore the rules (the Reasoning Anchors) and just use its own brain.

• With Rules: The AI was 99.7% accurate.
• Without Rules: The AI dropped to 96.5% accuracy, and its mistakes started spreading to other parts of the answer.

The Analogy: Think of the AI like a brilliant student taking a test. If you give them a blank sheet of paper, they might get distracted or daydream. But if you give them a checklist (the Reasoning Anchors), they focus perfectly and get the right answer every time. The checklist didn't make them smarter; it just kept them on track.

5. Why Does This Matter?

This isn't about replacing human experts. It's about giving them a super-powered assistant.

• Consistency: The AI never gets tired, never has a bad day, and always follows the rules.
• Safety: By understanding exactly what drivers were doing before the crash (were they distracted? were they speeding?), we can design better cars and safer roads to prevent the next crash.
• Scale: Humans can only look at a few hundred crashes a year. This AI can look at millions, finding patterns we might miss.

In a Nutshell:
This paper shows that by splitting a complex job into two steps and giving the AI a strict checklist to follow, we can build a system that reconstructs car crashes with perfect accuracy and lightning speed, helping us make the roads safer for everyone.

Technical Summary

1. Problem Statement

Traffic collision reconstruction traditionally relies on human experts to synthesize fragmented, multimodal evidence (witness statements, physical damage, scene diagrams, and Event Data Recorder (EDR) logs) to determine causal mechanisms and liability. While post-crash reconstruction is well-established, pre-crash reconstruction (inferring driver behavior and vehicle dynamics immediately preceding a collision) faces significant challenges:

• Data Fragmentation & Heterogeneity: Evidence comes from disparate sources (text narratives, tabular data, visual diagrams, time-series sensor data) that are often incomplete or contradictory.
• EDR Complexity: EDRs typically record only the final 5–10 seconds before a crash. In complex scenarios, a single collision may trigger multiple, overlapping, or ambiguous EDR records, making it difficult to identify the specific record corresponding to the "first harmful event."
• Human Limitations: Manual analysis is time-consuming, prone to cognitive bias, and suffers from inconsistency, especially when analysts lack specialized reconstruction training.
• LLM Limitations: While Large Language Models (LLMs) show promise, single-agent systems struggle with hallucination, lack of domain-specific reasoning, and difficulty in handling multi-step logical deduction under noisy data conditions.

2. Methodology

The authors propose a two-phase, multi-agent AI framework designed to reconstruct crash scenarios and infer pre-crash behaviors without fine-tuning (zero-shot evaluation).

A. Data Source & Preprocessing

• Dataset: 277 rear-end Lead Vehicle Deceleration (LVD) cases from the NHTSA Crash Investigation Sampling System (CISS, 2017–2022).
• Data Transformation: Raw multimodal data (PDF scene diagrams, text reports, structured EDR tables) were converted into a unified structured natural language format. This involved:
  • Crash Scene Description: Summarizing vehicle types, damage, and environmental factors.
  • EDR Data Report: Converting time-series sensor data into sequential textual records.
  • Scene Diagrams: Retained as images for vision-capable models.

B. The Two-Phase Multi-Agent Architecture

The framework utilizes specialized agents for distinct tasks, leveraging different types of LLMs:

• Phase I: Crash Reconstruction (Multimodal Understanding)

  • Agent: General-purpose multimodal LLM (selected: Claude 3.7).
  • Input: Scene diagram (image) + Crash Scene Description (text).
  • Task: Reconstruct the accident scene, identify vehicle spatial relationships, and generate a structured narrative of the collision sequence.
  • Mechanism: Uses a "Calibration Protocol" to prioritize text-based evidence over visual noise in diagrams.

• Phase II: First Crash Inference (Logical Reasoning)

  • Agent: Reasoning-centric LLMs (tested: DeepSeek-R1, Grok 3-mini, Gemini 2.5 Pro).
  • Input: Phase I Reconstruction Report + EDR Data Report.
  • Task: Identify the "striking" and "struck" vehicles in the first harmful event and pinpoint the most relevant EDR record(s) for each.
  • Key Innovation: Reasoning Anchors: The prompt includes five structured rules (Anchors) to constrain the model:
    1. Primary Understanding: Ground analysis in the Phase I reconstruction.
    2. EDR Filtering: Explicitly discard unrelated or "unknown" events.
    3. Missing Data Handling: Fallback to reconstruction reports if EDR is missing, avoiding fabrication.
    4. Critical Timing: Correctly interpret T=0 (trigger time) vs. actual collision time.
    5. Event Selection: Resolve overlapping EDR records by selecting the one closest to the first harmful event.

C. Evaluation Strategy

• Ground Truth: Established via consensus between two independent research analysts (non-specialists in reconstruction) for 39 "Complex" cases (many-to-one EDR mapping). A third-party LLM was used only to flag potential conflicts.
• Metrics: Accuracy in identifying striking/struck vehicles and the correct EDR event numbers.

3. Key Contributions

1. Novel Multi-Agent Framework: A two-stage architecture that separates multimodal perception (Phase I) from deep logical reasoning (Phase II), optimizing the use of different LLM types.
2. Structured Reasoning Anchors: The introduction of explicit, rule-based constraints within prompts to mitigate hallucinations and guide the model through complex temporal reasoning.
3. Zero-Shot Generalization: Demonstrated that high performance can be achieved without domain-specific fine-tuning, relying instead on prompt engineering and architectural design.
4. Scalable Pre-Crash Analysis: A system capable of processing fragmented, noisy data to isolate critical pre-crash time windows, enabling large-scale behavioral analysis.

4. Results

• Overall Accuracy: The framework achieved 100% accuracy across all 4,155 experimental trials (277 cases × 5 trials × 3 models). It correctly identified vehicle roles and EDR events in both simple (one-to-one mapping) and complex (many-to-one mapping) scenarios.
• Human vs. AI Comparison:
  • Simple Cases: Both AI and human analysts achieved 100% accuracy.
  • Complex Cases: Human analysts achieved 92.31% accuracy, while the AI framework maintained 100%.
• Cross-Model Consistency: Three distinct reasoning models (DeepSeek-R1, Grok 3, Gemini 2.5) produced identical outputs for every trial, proving that performance is driven by the structured prompt design rather than specific model capabilities.
• Ablation Study (Impact of Reasoning Anchors):
  • With Anchors: 99.7% case-level accuracy; errors were isolated to a single output type.
  • Without Anchors: Accuracy dropped to 96.5%; errors spread across multiple dimensions (vehicle roles and EDR events), indicating a collapse in reasoning stability.
  • Token Usage: Anchors increased reasoning-stage token consumption by ~38.7% but improved stability significantly.
• Efficiency: The AI framework processed cases in <1 minute (average 22–44 seconds depending on the model), whereas human analysts took an average of 6.47 minutes per complex case.

5. Significance

• Decision Support for Non-Specialists: The system provides a reliable, reproducible "first-pass" analysis for investigators who may not be certified crash reconstruction experts, reducing cognitive load and increasing consistency.
• Mitigation of Hallucination: By embedding "Reasoning Anchors," the study demonstrates a practical method to control LLM hallucinations in high-stakes, safety-critical domains.
• Foundation for Large-Scale Behavioral Research: By automating the identification of the correct pre-crash time window, this framework enables the extraction of driver behavior patterns (e.g., braking onset, deceleration profiles) from thousands of cases, which was previously infeasible due to manual processing costs.
• Future Applications: The structured textual reconstructions generated by the system can serve as inputs for emerging text-to-image/video technologies, facilitating visual accident reconstruction for education and legal proceedings.

In conclusion, this paper establishes that a carefully designed multi-agent architecture with structured reasoning constraints can outperform human analysts in complex, ambiguous traffic crash scenarios, offering a scalable and efficient solution for pre-crash analysis.