Work zones remain critical hotspots for crashes due to rapidly changing geometries, temporary traffic control, and heterogeneous data sources. Traditional safety modeling approaches often rely on centralized data, which limits scalability and raises privacy concerns. Federated learning (FL) offers a promising alternative by enabling decentralized model training while preserving data security. This study investigates FL as a decentralized modeling paradigm for dynamic work zone safety. The primary objective is to operationalize an FL-based framework capable of leveraging heterogeneous transportation data while preserving data locality. The proposed framework is demonstrated through proof-of-concept implementations involving three work zone-relevant components: visual perception of work zone elements, driver behavior modeling using naturalistic trajectory data, and contextual awareness derived from structured work zone information. For the perception component, FL is implemented across decentralized visual data sources to illustrate node-level integration through federated aggregation, while the behavioral and contextual components are evaluated using conventional modeling approaches for framework compatibility. Results indicate that FL can achieve performance comparable to or exceeding centralized baselines while improving scalability, modularity, and privacy preservation. The findings highlight FL as a viable foundation for adaptive and privacy-aware work zone safety modeling.
The resilience of urban roadway network is crucial for maintaining effective mobility during disruptions. However, traffic safety has received limited attentions in the domain of resilience assessment. The study introduces the concept of safety resilience, which integrates traffic safety consideration into resilience research. Unlike traditional resilience that primarily focuses on traffic efficiency metrics, safety resilience is defined as a network ability to maintain and restore safety performance under disturbances by evaluating real-time traffic risk. The study presents the evaluation of safety resilience based on traffic conflicts at intersections and on roadways and explores the impact of different traffic control measures, including lane control, signal control, and connectivity improvement, on both traffic safety and efficiency. The results demonstrate that the proposed Safety Resilience Index (SRI) provides an effective means to assess network safety resilience by capturing the trade-off between safety and efficiency. Lane restrictions are found to be more suitable for road segments with high safety risk, while lane widenings are more effective in mitigating severe congestion. Signal control can enhance safety resilience but increase travel time and rear-end conflicts on roadways. Although connectivity improvement can theoretically enhance resilience from the topological perspective, it may result in worse overall network safety and efficiency. These findings underscore the need for a balanced approach that integrates both safety and efficiency in enhancing urban network resilience.
The rapid growth of road freight has led to an increasing number of trucks, where their long braking distance, large mass, and limited maneuverability make them particularly vulnerable to safety risks, especially at high-risk nodes such as tunnels, bridges, and step grades. These risks are generated at the microscopic level but accumulate and propagate along the entire trip, creating the need for a modeling framework capable of capturing both local safety mechanisms and system-level traffic evolution. To address this need, this study proposes a multiscale hybrid traffic flow model that integrates an improved Cellular Automata (CA) model for detailed characterization of accident-prone segments and a Cell Transmission Model (CTM) for long-distance freight transport. Two transition areas are designed to ensure consistent information exchange between the discrete CA and continuous CTM domains. Using naturalistic driving data for calibration and validation, the hybrid model accurately reproduces truck-car interaction patterns and system-level traffic dynamics. The results show that vehicle heterogeneity and different vehicle combinations lead to distinct car-following behaviors and safety outcomes, as reflected in surrogate safety measures such as Time Headway and Time-to-Collision. Although trucks tend to adopt more cautious strategies, truck-truck interactions still exhibit the highest risk levels due to reduced maneuverability and accumulated disturbance. These findings provide actionable insights for freight operations and road safety management, supporting the development of targeted regulations and truck-specific safety strategies at high-risk locations.
Arterial wrong-way crashes (AWWCs) pose a safety problem on Florida roadways, yet most wrong-way driving research focuses on limited-access facilities with less attention on arterials. This article applies a corridor approach to examine AWWCs in Central Florida, identify arterial WWD hotspots, and demonstrate the transferability of this approach from South Florida to Central Florida. Arterial corridors were defined using context classification and through lane count, resulting in 1,475 corridors with 2,522 AWWCs. Negative binomial and logistic regression models used corridor variables to identify factors influencing AWWC frequency and probability. Both models revealed that corridors in suburban and urban contexts, urban principal arterials, urban minor arterials, rural principal arterials, corridors with higher signalized intersection densities, and one-way corridors exhibited elevated AWWC risk, while corridors with higher vegetation median proportions demonstrated reduced risk. Using model predictions, 47 corridors were identified as high-risk hotspots by both models. These shared hotspots were typically long, high-volume corridors with many signalized intersections, low vegetation median proportions, and were concentrated in Orange and Hillsborough Counties. These findings confirm that the corridor approach is transferable beyond South Florida and can provide agencies with a practical tool for screening roadway networks and prioritizing corridors for proactive safety treatments.
At tunnel entrances and exits, the abrupt transition in cross-sectional and lighting conditions poses substantial safety risks, which are further exacerbated in spiral tunnels due to their coupled radius-slope design. This study proposes an improved artificial potential field-based risk modeling framework that integrates geometric alignment (radius and slope), lighting conditions, driver characteristics, and vehicle dynamics, enabling more precise quantification of driving risk under the combined influence of multiple factors. Using three alignment design schemes from a real-world engineering project as prototypes for the spiral tunnel, driving simulation experiments were conducted to systematically quantify and compare driving risks at the spiral tunnel portals across different radius-slope combinations. Results show that radius and longitudinal slope near tunnel portals significantly affect driving risk. A small radius coupled with a steep slope markedly increases entrance risk. Compared with open-road sections, internal tunnel zones exhibit a higher risk, and the 100 m before the exit generally presents a greater risk than the 100 m after the entrance. This study clarifies the mechanisms underlying risks at spiral tunnel portals under complex alignment and environmental conditions, and provides theoretical and practical guidance for improving alignment design and mitigating portal risks.
In the context of China's new-type urbanization, rural roads have become high-risk traffic environments due to land-use changes, mixed traffic composition, and lagging infrastructure. Using crash data from five rural roads in Beijing during 2023-2024, this study constructed meso-level units by integrating macro-level urbanization characteristics with micro-level road attributes. Urbanization level was introduced as a moderating variable, and a negative binomial-conditional autoregressive (NB-CAR) model incorporating interaction terms was developed to investigate the factors influencing traffic crash frequency and the underlying mechanisms in urban-rural transitional areas. The results show that, compared with low urbanizing areas, moderately urbanizing areas are significantly positively associated with crash frequency, whereas highly urbanizing areas are significantly negatively associated with crash frequency. In addition, the interaction effects between the presence of central medians and urbanization level, as well as between segment length and highly urbanizing areas, indicate that crash risk arises from the combined influence of multiple factors rather than from any single factor. The study further suggests that traffic safety infrastructure and management capacity in moderately urbanizing areas often lag behind rapid traffic growth. These areas should therefore be prioritized in proactive risk governance, with systematic crash reduction pursued through fault-tolerant design and shared safety governance.
With the rise of intelligent driving and human-machine codriving, effective yet minimally intrusive risk prompts have become crucial for road safety. While important groundwork on prompt effectiveness has been laid in prior work, there is still an opportunity to build on this foundation and further investigate the joint parametric optimization of frequency, duration, and lead time. In extending prior efforts in scene classification, a need has emerged for the development of data-driven, quantitative approaches to urgency quantification that complement the existing methods. We conducted a within-subject driving simulator study with 40 drivers across nine risk scenarios. Using baseline physiological and behavioral data from a no-prompt control, we constructed an objective urgency index (UI) and evaluated how prompt-parameter combinations affected driving, eye movements, and physiology. The P2-D1-L3 setting (two prompts, 1 s duration, 3 s lead time) was Pareto-optimal, as it improved lateral stability while preserving longitudinal control. Lead time was the dominant factor, and it outperformed frequency and duration. Prompting produced the greatest benefits in medium-urgency scenes, taking an inverted U-shaped moderation effect. Mediation analysis indicated that improved lateral control was driven primarily by physiological arousal, which was indexed by pupil dilation. These findings quantify the optimal parameter boundaries and support the fine-grained design of risk-promotion HMIs in intelligent vehicles.
Roundabouts, serving as critical nodes in urban transportation networks, the analysis of conflicts between motorized vehicles (MVs) and non-motorized vehicles (NMVs) is complex. Existing research has failed to adequately consider the impact of trajectory characteristics on conflict analysis modeling and has ignored the multidimensional mapping relationship between conflict levels and micro-level traffic behavior, resulting in the simplification of traditional conflict indicator classification. Therefore, this study introduced Circular Trajectory Entropy (CTE), Radial Deviation Index (RDI) and Curvature Change Rate (CCR) to measure the randomness of MV and NMV trajectories, and used a multi-class logistic model to quantify the impact of micro-characteristics and trajectory fluctuations on the severity of conflicts. Statistical results showed that MVs mainly adopted conservative deceleration behavior in response to potential conflicts, whereas NMVs exhibited higher path randomness and greater trajectory fluctuations at roundabouts. A logistic regression model based on multi-feature fusion performs exceptionally well in analyzing minor and severe conflicts. When using SHAP plots to reveal key influencing factors, it was found that CCR was the most influential feature in distinguishing between minor and severe conflicts. These findings can be used to evaluate the severity of conflicts between MVs and NMVs at signal-free roundabouts.
According to the World Health Organization, road traffic accidents contribute to approximately 1.3 million deaths annually, posing a global public safety challenge. Current deployment strategies for traffic police face difficulties in balancing accident prevention and real-time response, leading to inefficiencies. This study proposes a two-phase framework for traffic police dispatch, integrating pre-deployment and dynamic emergency response. In the pre-deployment phase, a deep learning model using accident data from Yinzhou District, Ningbo City (April 2020 to October 2021) identifies high-risk zones-defined as statistically significant spatial clusters of high accident frequency-for police positioning. The dynamic dispatch phase utilizes a real-time optimization algorithm based on the golden hour (<= 10 min) to reduce response times. Experimental results show a 28.71% improvement in average response time (from 5.54-7.21 min to 4.52 min) and an 85.43% police utilization rate, outperforming baseline methods. This approach improves response efficiency and enhances officer utilization. The framework demonstrates potential scalability, offering insights for improving road safety in urban settings with similar characteristics.
A deep understanding of short-term behavioral patterns is critical for predicting stop/go behavior, especially in complex scenarios such as approaching signalized intersections demanding stop/go decisions. Although driving behavior has been studied from static and dynamic perspectives, short-term behavioral-pattern changes during signal phase transitions at intersections and their contribution to stop/go prediction remain underexplored. This study proposes a stop/go prediction model integrating short-term behavioral-pattern identification using UAV-collected trajectories from urban intersections. A multi-type feature space is first constructed from continuous trajectories to represent short-term behavior. Based on this representation, the trajectories are segmented into driving behavior primitives using Bayesian Model-based Agglomerative Sequence Segmentation (BMASS). The resulting primitive segments are then grouped into discrete primitive types using a Gaussian Mixture Model (GMM), from which primitive-label sequences are obtained. Based on these sequences, Latent Dirichlet Allocation (LDA) extracts short-term behavioral-pattern proportions, while a first-order Markov chain captures behavioral-pattern transition statistics. Finally, these behavioral-pattern features are fused with trajectory-based statistical features and fed into an XGBoost model for stop/go prediction. Compared with the trajectory-only baseline, the full framework improves accuracy from 0.913 to 0.971. The results indicate that short-term behavioral-pattern features offer incremental predictive value beyond trajectory- and primitive-level information.
Autonomous vehicles (AVs) are expected to significantly enhance transportation safety in the coming decades. As the development and deployment of AVs continue to grow, it is crucial to understand the risks and factors contributing to collisions, particularly across different operational modes. Most prior studies aggregate AV crash records without distinguishing whether or not the vehicle was operating autonomously or manually at the time of the crash, limiting insights into mode-specific crash dynamics. This study addresses this gap by using crash records from the California Division of Motor Vehicles, statistically segmented by operational mode and analyzed with advanced statistical methods. Key findings include: (1) rear-end collisions are the most common crash type across both operational modes; (2) autonomously operated AVs are more likely to be involved in severe crashes under poor lighting conditions; (3) head-on collisions significantly increase the likelihood of severe outcomes for autonomously operated AVs, even during daylight conditions; and (4) on average, the probability of a severe crash is 25.70% lower when a vehicle is operating autonomously compared to manual control. The implications of these findings are extensively discussed to provide valuable guidance for transportation policymakers and researchers in addressing AV deployment and developing effective strategies to mitigate collision risks in mixed-fleet conditions.
Traffic safety remains a critical priority within urban transportation systems, and real-time safety analysis is essential for timely risk identification and proactive intervention. Most existing studies on conflict prediction have examined the effects of traffic conditions, vehicle status, and geometric factors on conflict risk. However, driver behaviors, especially conflict evasive behaviors, also play a crucial role. To address this gap, this study analyzed vehicle evasive patterns during rear-end conflicts. An improved K-Means clustering algorithm was applied to identify six evasive patterns under both high- and low-risk conflict conditions. Key influencing factors included roadway and traffic characteristics, following- and leading-vehicle attributes, and non-conflict driving behaviors under normal driving conditions. Following variable selection, a two-level Nested Logit (NL) model was established to jointly predict evasive patterns and conflict risk levels, achieving predictive accuracy above 0.70 and AUC greater than 0.90. Key findings revealed that braking behavior under non-conflict conditions was strongly correlated with subsequent evasive responses, whereas congested traffic conditions tended to delay driver reactions. Leading-vehicle characteristics, such as speed, acceleration, and headway, significantly influenced following-vehicle responses. These findings suggest that incorporating evasive driving patterns into conflict risk prediction can provide a more comprehensive and individualized understanding of rear-end conflicts on expressways.
The commercial viability of aqueous zinc metal batteries is bottlenecked by sluggish Zn-2(+) desolvation kinetics and erratic interfacial charge transfer. Here, we report a catalytic strategy using 4-dimethylaminopyridine (DMAP) as an interfacial molecular catalyst that undergoes structural interconversion between neutral phenolic and quinoid resonance forms. Unlike passive interfaces, this pi-conjugated molecule orients at the metal-electrolyte interface to function as a desolvation relay. Through electron delocalization and reversible lone-pair coordination, the active interface transiently replaces water ligands, lowering the activation energy for Zn-2(+) reduction. This catalytic interconversion facilitates the in situ formation of a vertically ordered, ion-conductive interphase that homogenizes Zn-2(+) flux and eliminates parasitic reactions. As a result, Zn||Zn symmetric cells operate stably for 1500 h at 1 mA cm(-)(2) and 300 h at 20 mA cm(-)(2) with 10 mAh cm(-)(2), while Zn||Cu cells exhibit highly reversible plating/stripping for over 1800 cycles. Full Zn||Br-2 cells deliver 225 mAh g(-)(1) over 1500 cycles with Coulombic efficiency exceeding 99.25%. Our findings demonstrate that transitioning from passive to catalytic interface design is a transformative route for high-rate, long-life aqueous energy storage.
Traffic conflicts are commonly used as surrogate safety measures for intersection safety evaluation. The objective of this study is to conduct a systematic cross-comparison of full Bayesian approaches for traffic conflict modeling. Five full Bayesian models were developed and compared, including Poisson lognormal (PLN) model, random intercept PLN (RI-PLN), random parameters PLN (RP-PLN), spatial PLN (S-PLN), and temporal PLN (T-PLN) models. The extended PLN models aim to account for unobserved heterogeneity and spatial/temporal correlation among traffic conflicts. The five models are estimated using traffic conflict data extracted from unmanned aerial vehicle-based vehicle trajectories collected at 17 urban intersections in Athens, Greece. Model performance is evaluated using the Deviance Information Criterion and posterior diagnostics. Results show that the T-PLN model provides the best goodness-of-fit, indicating the importance of temporal correlation in modeling traffic conflicts. Incorporating temporal correlation yields greater performance than accounting for spatial correlation or random parameters alone. Parameter estimates from the favorite model indicate that taxis, motorcycles, heavy vehicles, and signalized control are significantly associated with traffic conflicts. These findings highlight the importance of explicitly modeling temporal correlation in conflict-based safety analysis and offer practical insights for intersection safety management and signal operation optimization.
Conditionally automated driving vehicles (CADVs) have been deployed in real-world testing and operational scenarios. However, research into their safety impacts on mixed traffic flow dynamics remains inadequate. Critical gaps persist in identifying key risk factors and understanding the mechanistic effects of various predictors. To address these research gaps, this study investigates the safety implications of CADVs during the transition to manual driving mode in mixed traffic environments. Vehicle trajectory data were collected through driving simulator experiments and used to calibrate parameters of a car-following model. Microscopic simulation tests were then designed and conducted, incorporating factors such as time budget, CADV penetration rate, takeover duration, and traffic volume. Based on the collected trajectories, time to collision was employed to evaluate rear-end conflict risks between the ego vehicle and the leading vehicle. An explainable machine learning model was developed to predict and interpret the impact of various factors on rear-end conflicts resulting from control transitions in CADVs. The model identified the five most influential factors affecting the ego vehicle's rear-end conflict risk: acceleration and type of the leading vehicle, speed and following distance of the ego vehicle, and the speed difference between the two. The risk of rear-end collision was found to be negatively correlated with the ego vehicle's speed, following distance, and the leading vehicle's acceleration, whereas a greater speed difference increased conflict risk. Furthermore, control transitions in CADVs not only compromise their own safety but also increase the rear-end collision risk for following vehicles. These findings clarify the key factors influencing rear-end conflict risks during control transitions and highlight the broader safety implications of CADVs in mixed traffic, providing a basis for enhancing control strategies in such scenarios.
This study disaggregates pedestrian injury severity outcomes by both vehicle type and vehicle generation to statistically examine how pedestrian risk has evolved in the modern U.S. vehicle fleet. Using nationally representative Crash Report Sampling System data from 2021 to 2023, random-parameter logit models with heterogeneity in means are estimated separately for passenger cars and light trucks (which includes SUVs, and pick-up trucks) across three model-year cohorts: older vehicles manufactured between 2000 and 2010, transitional vehicles produced between 2011 and 2015, and modern vehicles produced in 2016 or later. This disaggregation reflects major shifts in vehicle mass, geometry, and the diffusion of pedestrian-related safety technologies over the past two decades. The results reveal substantial generational instability in pedestrian injury-severity relationships, particularly for light-truck vehicles. Out-of-sample simulations show that, under identical crash circumstances, transitional and modern vehicles in both classes are associated with higher predicted probabilities of severe pedestrian injury compared to older vehicles. While light trucks exhibit higher severe-injury risk than passenger cars in older generations, this class-based gap narrows substantially in modern vehicles, indicating a convergence in pedestrian injury severity outcomes. These findings demonstrate that treating vehicle classes as static masks important generational effects and highlight the need to explicitly account for vehicle generation when evaluating pedestrian risk and developing safety countermeasures.
The complicated driving environment near the tunnel entrance increases the risk of lane-changing due to variations in speed limits and lighting environment. Therefore, the objective of this study is to investigate the factors affecting lane-changing risk in the transition segment of tunnel entrance using vehicle trajectory data collected by naturalistic driving tests. To assess the level of lane-changing risk and explore the effects of different traffic and environmental factors on lane-changing risk, the conflict index for lane-changing (CILC) and the complexity of lane-changing environment are defined. Four random parameter ordered logit (RPOL) models with heterogeneity in means and variances are developed based on four patterns of complexity partitioning, which are compared with the general RPOL models, showing that the former has better performance. Model estimation results reveal that several variables significantly affect the risk as well as unobserved heterogeneity. The results also indicate that the varying complexity of lane-changing may cause the same variables to have opposite effects on risk, such as the lateral acceleration and lateral velocity of the SV. The interpretative findings provide valuable information for the analysis of lane-changing risk and the design of Advanced Driver Assistance Systems.
As urban road networks become increasingly complex, managing safety is a critical challenge. Proactive safety management requires both accurate risk prediction and an understanding of contributing factors. This study introduces a Temporal Fusion Transformer (TFT)-based framework that integrates prediction, interpretation, and intervention analysis for large-scale safety improvement. The TFT predicts a risk index for each intersection and link, utilizing both static and dynamic variables. Its dynamic attention mechanism provides insights into how the effect of each variable shifts over time. Trained and validated on the high-resolution pNEUMA trajectory dataset from Athens, Greece, the model achieved strong predictive performance (MAE = 0.30, RMSE = 0.461). Temporal analysis of attention weights revealed that taxis and motorcycles consistently contribute the most to elevated safety risks, while medium vehicles were most influential during the early morning hours. These insights enable the development of targeted strategies, such as dynamic access restrictions or time-based charges for specific vehicle classes during high-risk periods. A sensitivity analysis further shows how modest changes in traffic composition can significantly improve safety, quantifying the benefits of proposed interventions. Overall, the framework offers transportation agencies a practical, data-driven tool that bridges the gap between advanced predictive modeling and real-world safety management strategies.
To address safety hazards at unsignalized intersections on national highways, this study proposes a probability-consequence-resolution (PCR) three-dimensional indicator system for traffic conflict classification and develops a multi-intersection joint safety assessment framework. Using road users' trajectory data and macroscopic intersection features from four typical unsignalized intersections on national highways in Hebei Province, the study applies the proposed three-dimensional metrics in conjunction with the K-means clustering method to classify conflict severity levels. The persistence and evolution patterns of traffic conflicts were subsequently investigated. A safety assessment framework was established for individual conflicts and holistic intersection assessments, followed by XGBoost machine learning modeling to identify influencing factors. Key findings demonstrate: (1) Traffic conflicts were classified into four severity levels with dynamically adjustable classification thresholds; (2) Continuity was observed in conflict processes; (3) Conflict severity was primarily influenced by the overlap time of conflicting road users; (4) Both conflict count and intersection safety scores exhibited significant positive correlations with violating road users and traffic volume; (5) Among macroscopic features, road surface defects and the number of main-road lanes were identified as the dominant predictors. This methodology provides a new paradigm for safety diagnostics at unsignalized intersections.
Red-light running (RLR) remains a significant traffic safety concern at signalized intersections, contributing to severe crashes and endangering road users. Despite its significant safety implications, research on factors contributing to injury severity in RLR-related crashes remains limited. This study applied Association Rule Mining (ARM) to analyze severity patterns of RLR-related crashes at signalized intersections in Arizona from 2013 to 2024. Using the Apriori algorithm, strong associations among crash factors were identified based on support, confidence, and lift metrics. The analysis found that the combination of impaired driving, distracted driving, and right-angle collisions significantly contributed to injury crashes. Subsequently, the study separately examined crash severity patterns for severe and minor injury crashes to uncover distinct contributing factors. The generated rules revealed that severe injury crashes (i.e. K and A crashes on the KABCO scale) were strongly associated with the co-occurrence of male drivers, non-motorist involvement, impaired driving, distracted driving, speeding, nighttime conditions, and undivided roadways. Conversely, minor injury crashes (i.e. B and C crashes on the KABCO scale) were more commonly associated with the co-occurrence of middle-aged and female drivers, weekday, and dry road conditions. These insights offer valuable data-driven strategies for transportation agencies and policymakers to enhance intersection safety.
