Holistically Identifying Road Complexity and Relating It to Fatal Crashes

Wang, Meng; Roberts, Shannon C · 2025 · ROSA P / New England University Transportation Center

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This study addresses the need to holistically understand roadway complexity and its relationship to fatal crashes, aiming to improve traffic safety and advance automated driving systems. Previous research often relied on isolated data sources, such as imagery or behavioral metrics, without integrating them. The authors define "roadway complexity" as the combination of scene complexity (semantic and contextual features) and driving behavior (vehicle kinematics). To investigate this, they propose a two-stage framework that extracts hidden contextual information from these integrated features to predict crash likelihood. The methodology utilizes a subset of the MIT-AVT naturalistic driving dataset, comprising 500 video clips across various scenarios (highway, rural, urban, etc.) and 10,407 extracted frames. The study generates three types of features: semantic features derived from the OneFormer algorithm, kinematic features from CAN bus data (e.g., speed, acceleration), and contextual features generated by the GPT-4o Large Language Model (LLM). Ground truth data includes a complexity index annotated by both the LLM and human workers via Amazon Mechanical Turk, as well as crash density values calculated from Massachusetts crash data (2018–2022) using Kernel Density Estimation. The experimental design involves a complexity encoder (a neural network) that learns latent features from the input data, followed by a crash prediction model using algorithms like Random Forest and Neural Networks. The results demonstrate that integrating semantic, kinematic, and contextual features yields the best performance. The crash prediction model achieved an accuracy of 87.98% using original features alone, which improved to 90.46% when latent complexity features were added. Statistical tests confirmed these improvements were significant. Analysis using SHAP values revealed that vegetation area was a key differentiator for low-crash-density areas, while speed-related features were most influential for high-crash-density areas, with lower speeds associated with higher crash density. Notably, complexity indices generated by the LLM outperformed those from human annotators in predicting crash rates, suggesting AI-based tools offer superior scalability and accuracy for this task. The significance of this work lies in its demonstration that a holistic integration of scene and behavioral data significantly enhances crash prediction accuracy. The findings support the development of advanced driver assistance systems and driver monitoring systems by providing a robust method for assessing environmental risk. Furthermore, the framework offers highway engineering departments data-driven insights to identify and mitigate risk factors in crash hotspots, aligning with the U.S. Department of Transportation’s Safe System Approach. This proactive intervention potential could lead to reduced roadway fatalities and optimized infrastructure investments.

Key finding

Integrating latent complexity features derived from semantic, kinematic, and contextual data into crash prediction models significantly improves accuracy, with large language model-generated complexity indices outperforming human annotations.

Methodology

naturalistic

Sample size: 500

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction
• road complexity
• incidence prevalence
• naturalistic crash near crash
• urban rural setting
• pre crash contributing factors

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: crash risk outcomes
• Methodological Resource: dataset resource
• Theoretical Contribution: computational model