Modeling Drivers' Risk Perception via Attention to Improve Driving Assistance

Biswas, Abhijat; Gideon, John; Tamura, Kimimasa; Rosman, Guy · 2024 · arXiv

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 paper addresses the high rate of superfluous alerts in Forward Collision Warning (FCW) systems, which often ignore driver awareness and cause alert fatigue. The authors propose modeling drivers’ risk perception by incorporating their attention to specific traffic objects. They assume that during inattentive periods, drivers extrapolate other vehicles’ motion using a constant velocity counterfactual, leading to inaccurate risk estimates. The study utilizes the AVT-FCW-TrajAttn dataset, derived from real-world FCW deployments. The authors extracted 3D vehicle trajectories using Visual SLAM and monocular 3D detection, and annotated driver gaze-to-objects and FCW validity from video recordings. They developed two approaches: a learned trajectory forecasting model using a Planning Transformer with counterfactual inputs, and an attention-aware conventional FCW algorithm that adjusts warning distances based on the difference between actual and counterfactual lead vehicle velocities. Experiments on 60 episodes showed that the attention-aware conventional FCW method outperformed baselines, achieving a True Positive Rate of 0.733 and True Negative Rate of 0.710. In contrast, learned models suffered from high false positive rates due to inconsistent joint trajectory predictions, while standard conventional FCW had low detection rates. The proposed method significantly reduced false positives and improved overall accuracy compared to gaze-only or traffic-unaware models. The findings demonstrate that integrating driver attention into risk estimation improves FCW performance by reducing nuisance alerts. This approach mitigates alert fatigue and enhances on-road safety by ensuring warnings align with the driver’s actual perceptual state and scene awareness.

Key finding

Machine learning models successfully capture individual differences in risk perception, with drivers showing consistent but distinct risk assessment patterns across scenarios that can be modeled with relatively few parameters.

Methodology

lab_experiment

Sample size: 58

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 discover_arxiv on 2026-05-04 (4 acquisition events logged).

StageOutcomeToolModelPromptAttemptsCompleted
discover success arxiv 3 2026-05-04
archive success 1 2026-05-04
extract success cached 2 2026-06-07
clean success 1 2026-06-01
chunk success 1 2026-06-01
embed success 1 2026-06-02
enrich success 1 2026-05-04
promote success 1 2026-05-04
summarize success llm qwen3.6-27b-prismaquant summ-v5 2 2026-06-07
tag success vector_similarity 17 2026-06-11
verify success 1 2026-05-08

Summary generated by qwen3.6-27b-prismaquant on 2026-06-07; 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).