Predicting Glare-Related Traffic Outcomes with Transformer-Based Explainable Tabular Deep Learning

Engineering notes

Key topics: autonomous driving, deployment. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract Glare from both low sun angles under daytime conditions and headlight exposure under nighttime conditions degrades driver visibility and remains a persistent challenge for safe roadway operation, yet the mechanisms governing outcome differentiation under glare conditions across complex roadway, vehicle, and driver interactions remain insufficiently understood. Prior studies have primarily relied on econometric or conventional machine learning approaches, limiting their ability to capture nonlinear relationships and deliver transparent, decision-oriented insights. This study analyzes 18,073 glare-related traffic records from Texas from 2017 to 2024 and evaluates four tabular deep learning architectures (TabR, MITRA, TabTransformer, and ARM-Net) to model three discrete outcome levels under glare conditions. A hybrid SMOTEENN resampling strategy is applied to address class imbalance, and SHapley Additive exPlanations (SHAP) are used to provide class-level and instance-level interpretability. Results show that transformer-based tabular models consistently outperform attention-based and residual baselines in predictive performance and outcome separability, while explainability analysis highlights restraint use, airbag deployment status, lighting environment, roadway context, speed regime, and driver demographic characteristics as the dominant contributors to outcome differentiation. Comparison with tree-based baselines further indicates that, although ensemble methods achieve competitive overall accuracy, the leading tabular deep learning architectures offer stronger sensitivity to fatal crash outcomes and support instance-level explainability through SHAP. These findings demonstrate the value of combining tabular deep learning with explainable artificial intelligence to support transparent, scalable modeling of glare-affected traffic outcomes and to inform targeted, data-driven strategies for mitigating glare-related transportation risks.

Links and sources

• PDF from original source

Comments