Vehicle–pedestrian crashes with similar predicted fatal-crash risk may arise from different combinations of contributing conditions, while the distinction between risk magnitude and risk composition has received limited attention. This study investigated whether reproducible multivariable contribution architectures exist among vehicle–pedestrian crashes with comparable model-predicted fatal-crash risk. Nationwide police-reported crash data from Japan were analyzed using temporally separated 2023 and 2024 cohorts, comprising 74,002 vehicle–pedestrian crashes, including 1,838 fatal crashes. Four classification models were evaluated using the 2023 cohort, with the 2024 cohort reserved for independent temporal validation. XGBoost was selected as the nonlinear base model and achieved a PR-AUC of 0.203 and ROC-AUC of 0.877 in the 2024 test cohort. SHapley Additive exPlanations (SHAP) were then used to decompose model predictions, and signed SHAP contributions were conditioned on overall predicted risk, aggregated across seven domains, normalized, and clustered. Two contribution profiles were identified. One showed relatively greater contributions from road-user characteristics, whereas the other showed greater contributions from environmental conditions, roadway geometry, and traffic control and speed regulation. The profiles were highly reproducible in 2024, with high centroid cosine similarities and an adjusted Rand index of 0.975 relative to an independently fitted 2024 solution. Both profiles occurred across the full predicted-risk distribution and were not well reproduced by simple partitions based on predicted risk, pedestrian age, or day/night conditions. Profile-specific SHAP interaction analysis further revealed systematic differences in nonlinear predictor combinations, particularly the interaction between pedestrian age and day/night conditions. These findings demonstrate that similar predicted fatal-crash risk can arise through distinct and temporally reproducible contribution architectures, suggesting that predicted risk magnitude alone does not fully characterize the multivariable circumstances underlying fatal vehicle–pedestrian crashes.
Coordinating multiple traffic interventions does not necessarily lead to safer or more efficient traffic operations, because redundant or poorly timed interventions may introduce additional disturbances. This study moves beyond the conventional focus on how multiple interventions should cooperate by formulating coordination as a hybrid decision-making problem that determines whether additional interventions are needed, which interventions to activate, and how to implement them. Accordingly, an on-demand coordination method is proposed to selectively coordinate adaptive traffic signal control, dynamic speed regulation, and route guidance in emerging mixed traffic comprising human-driven vehicles and connected and automated vehicles (CAVs). The method employs two functionally differentiated deep reinforcement learning agents to decouple primary signal control from auxiliary intervention selection while jointly optimizing safety and efficiency. To support risk-aware decision-making, extreme value theory (EVT) is integrated into the control framework to characterize approach-specific tail crash risk. Specifically, a non-stationary generalized extreme value model is developed to capture extreme traffic conflicts and derive value-at-risk estimates, enabling the controller to identify and prioritize severe risk states that may be overlooked by average or instantaneous safety measures. The proposed method is evaluated at two signalized intersections through simulations across various CAV market penetration rates. Compared with non-coordinated and coordinated benchmarks, the proposed method consistently reduces average waiting time and generally reduces upper-tail crash risk. These safety and efficiency benefits are accompanied by a moderate increase in fuel consumption, with the trade-off becoming less pronounced as CAV penetration increases. The findings of this study support a shift from indiscriminate multi-intervention control toward selective coordination and highlight the potential to extend EVT-based safety analysis from post hoc risk assessment to active control decision-making.
Collapsible bicycle helmets address portability as a barrier to helmet use, particularly for bicycle share scheme users, but whether they provide equivalent protection to conventional helmets is unknown. We determined whether head injury risk differs between designs and whether any difference is explained by mass and rotational protection technology.Thirty commercially available helmets, 24 conventional and 6 collapsible, were tested at 6.5 m/s onto a 45° anvil at four locations using an EN 17950:2024 headform. AIS4 + linear and AIS2 + rotational injury risks were estimated from peak linear acceleration and Brain Injury Criterion (BrIC). Linear mixed effects models (360 observations; design as fixed effect, helmet ID as random effect) assessed injury risk before and after adjustment for mass and rotational technology.Collapsible helmets had greater linear injury risk than conventional (0.242 ± 0.086 vs. 0.197 ± 0.051, p < 0.001), but this was non-significant after adjustment for mass and rotational protection. Rotational risk did not differ between designs. Collapsible helmets were heavier (461 ± 58 g vs. 294 ± 53 g), and mass predicted linear risk by + 0.002 per 10 g. Rotational technology reduced linear risk by − 0.026 and rotational risk by − 0.073 on average.Collapsible helmets exhibited greater linear injury risk than conventional designs, potentially through their greater mass and the absence of rotational protection technology. These findings guide safer collapsible designs that have the potential to facilitate helmet use and reduce head injuries, particularlyfor bicycle share scheme users.
Reinforcement learning (RL)-based adversarial scenario generation through dynamic agent–environment interactions has emerged as a promising approach for accelerating autonomous vehicle (AV) testing. However, existing RL approaches struggle to balance adversariality and naturalness because naturalistic driving behaviors avoid risk whereas adversarial behaviors intentionally induce it. The reward-driven nature of RL can further lead to objective polarization under linearly weighted rewards, resulting in scenarios biased toward a limited range of adversariality–naturalness trade-offs. This study proposes a preference-conditioned safety–critical scenario generation framework based on multi-policy multi-objective reinforcement learning (MORL), which models adversariality and naturalness as separate components of a vectorized reward and jointly optimizes them across the preference space. The adversariality reward incorporates risk–liability screening to filter two predefined invalid-risk patterns, while the naturalness penalty leverages a computationally efficient shadow-mode evaluation to provide real-time behavioral-deviation feedback relative to a reference driving policy. Envelope Q-learning is adopted to learn multiple preference-conditioned policies and is enhanced by a Kullback–Leibler divergence–based policy differentiation mechanism that mitigates policy mode collapse and encourages diverse behavioral evolution. Experiments in highway car-following and lane-changing environments show improvements in the evaluated scenario-quality and diversity metrics over the RL baselines. Compared with multi-round linearly weighted RL, the proposed method achieves 2.70 times the Pareto-front hypervolume, increases the valid safety–critical scenario success rate from 4.4% to 14.6%, and achieves 2.83 times the feature-space coverage volume. The results support the framework’s ability to generate diverse safety–critical scenarios across testing preferences within the studied highway operational design domain.
Unmanned aerial vehicle (UAV) safety is becoming increasingly critical as UAV operations are integrated into urban service and transportation systems, creating an urgent need for risk representations that can support refined low-altitude airspace management. Existing urban safety maps are still mainly expressed as two-dimensional or height stratified representations, which cannot capture the continuous variation of risk with altitude and therefore remain inadequate for continuous low-altitude flight scenarios. This paper develops a continuous three-dimensional spatiotemporal safety mapping framework for urban low-altitude environments. The framework integrates semantic information, digital surface model (DSM) and digital terrain model (DTM) derived morphology, pedestrian related exposure, and meteorological conditions within a unified geo-referenced support. A vertical continuous interpolation function is introduced to transform discrete risk representations into a continuous three-dimensional spatiotemporal risk field, thereby enabling continuous risk analysis across space, altitude, and time. Experiments in a real urban scene show that risk values vary significantly across different regions, time periods, and scene conditions, and that the resulting field reveals clear vertical differentiation and temporal redistribution patterns that conventional planar safety maps cannot represent consistently. These findings show that the proposed reusable continuous three-dimensional and time varying risk model offers a practical basis for refined low-altitude airspace management and has the potential to support future high density UAV operations.
