Mining patterns of autonomous vehicle crashes involving vulnerable road users to understand the associated factors
DOI: 10.1016/j.aap.2021.106473
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This study investigates the safety implications of autonomous vehicles (AVs) on vulnerable road users (VRUs), such as pedestrians, bicyclists, and scooterists, by analyzing crash patterns and associated factors. While AVs are expected to reduce crashes by eliminating human error, their specific impact on VRUs remains unclear due to a scarcity of empirical crash data. Previous research has largely relied on simulations, surveys, or virtual reality, often overlooking indirect VRU involvement—scenarios where an AV yields to a VRU but collides with another vehicle. To address this gap, the authors utilized text mining techniques on actual crash narratives to identify direct and indirect VRU involvement patterns. The researchers analyzed four years (2017–2020) of AV crash data collected from the California Department of Motor Vehicles, comprising 252 total crashes. The methodology involved two primary tasks: unsupervised text network analysis (TNA) to visualize keyword associations and supervised machine learning to classify crashes. For classification, the study compared four algorithms—Support Vector Machine (SVM), Naïve Bayes (NB), Random Forest (RF), and Neural Network (NN)—using metrics such as accuracy, precision, and F-1 score. To handle class imbalance, as VRU-involved crashes constituted only 14% of the dataset, the authors employed resampling techniques, finding that a combination of under-sampling and bootstrap yielded the best performance. The results identified 35 crashes involving VRUs: 22 direct involvements (collision between AV and VRU) and 13 indirect involvements. In direct crashes, bicyclists and scooterists were the primary participants, with bicyclists frequently at fault, often striking the AV’s rear bumper while the vehicle was stopped at a red light in autonomous mode. These crashes rarely resulted in injuries or police involvement. Conversely, indirect crashes predominantly involved pedestrians, occurring when the AV yielded to them and was subsequently hit by a conventional vehicle. These indirect incidents were more likely to result in minor damage to the AV’s rear bumper. Feature importance analysis from the best-performing classifiers (RF and NN) identified crosswalks, intersections, traffic signals, and specific AV movements (turning, slowing, stopping) as key predictors of VRU-related crashes. The study concludes that text mining of crash narratives is essential for capturing indirect VRU involvement, which traditional crash data often misses. The findings provide actionable insights for AV operators and city planners, highlighting that intersections and crosswalks are critical risk areas. Specifically, the data suggests that bicyclists and scooterists pose a higher risk of direct collision, often due to fault on their part, while pedestrians are more frequently involved in indirect scenarios where AV yielding behavior leads to rear-end collisions. These patterns underscore the need for refined AV programming and infrastructure planning to mitigate risks in mixed-traffic environments.
Key finding
Bicyclists and scooterists are more likely to be directly involved in autonomous vehicle crashes and are often at fault, whereas pedestrians are more frequently involved indirectly, particularly when the vehicle is in autonomous mode yielding right of way.
Methodology
dataset
Sample size: 252
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 author_sweep_intake on 2026-05-27.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-27 |
| archive | success | canonical_url | — | — | 7 | 2026-06-06 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | skipped | — | — | — | 3 | 2026-06-04 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
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.
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