Accident; analysis and prevention最新文献

{"title":"A knowledge-integrated learning framework for accurate quantification and semantic interpretation of driving aggressiveness","authors":"Zhaokun Chen ,&nbsp;Wenshuo Wang ,&nbsp;Chaopeng Zhang ,&nbsp;Yingqi Tan ,&nbsp;Lu Yang ,&nbsp;Junqiang Xi","doi":"10.1016/j.aap.2025.108225","DOIUrl":"10.1016/j.aap.2025.108225","url":null,"abstract":"<div><div>Aggressive driving is a major contributor to traffic fatalities, necessitating reliable assessment methods to guide driver interventions. Existing methods, however, lack granularity in assessing both the severity and specific maneuver categories of aggressive driving behaviors. This paper proposes a novel framework for multidimensional aggressiveness assessment using lateral-longitudinal acceleration and vehicle speed. The framework combines domain-specific prior knowledge with a non-parametric statistical method to quantify aggressiveness levels and automatically extract aggressive driving samples. We then classify them into distinct maneuver categories through fuzzy clustering and semantic analysis, assigning each sample a membership degree for every category. Finally, we integrate the samples’ levels with their membership distribution across the maneuvers to generate comprehensive profiles of individuals’ driving aggressiveness. Experimental validation with real-world driving data (<span><math><mrow><mi>N</mi><mo>=</mo><mn>90</mn></mrow></math></span> drivers) and real-time in-vehicle testing confirms our framework’s effectiveness and practicality. Additionally, a spatiotemporal analysis of driving maneuvers reveals insights into the evolution of aggressive driving and its relationship with environmental factors.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"222 ","pages":"Article 108225"},"PeriodicalIF":6.2,"publicationDate":"2025-09-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145004763","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Clear zone or W-beam guardrail for rural highways? A Full Bayes before-after evaluation by employing the Poisson Gamma-Lindley model","authors":"Shinthia Azmeri Khan,&nbsp;Hassan Bin Tahir,&nbsp;Shamsunnahar Yasmin,&nbsp;Shimul Md Mazharul Haque","doi":"10.1016/j.aap.2025.108218","DOIUrl":"10.1016/j.aap.2025.108218","url":null,"abstract":"<div><div>Run-off-road crashes often result in fatal and severe injuries on rural highways. Clear zones and W-beam guardrails are often installed to reduce run-off-road crashes and the extent of severe injuries. While the effectiveness of clear zone improvement and W-beams has been studied, their comparative effects in a similar empirical setting are not well known. At the same time, crashes on rural highways are often characterized by a high frequency of zero counts; however, models capable of addressing this issue—such as the Poisson-Gamma Lindley models—have not yet been tested in before-after evaluations of engineering treatments. Towards this end, this study aims to examine the safety effectiveness of clear zones and W-beam guardrails on rural highways by applying a Full Bayes approach with Poisson-Gamma Lindley model that considers heterogeneities resulting from the preponderance of zeros. A total of 82.70 km of treated roadway segments with Clear zone improvements (0–5 m to 5–10 m) implemented in 2014 and 2015 were extracted from Bruce Highway in Queensland, Australia. W-beam guardrails were installed along a total of 45.5 km of the same rural highway between 2013 and 2017. The computed Crash Modification Factors (CMFs) show that the clear zone widening significantly reduced total injury (CMF = 0.57), fatal, serious, and moderate injury (CMF = 0.58), fatal and serious injury (CMF = 0.53), and run-off-road crashes (CMF = 0.63). W-beam guardrails are also found to significantly reduce total injury (CMF = 0.75), fatal, serious, and moderate injury (CMF = 0.77), fatal and serious injury (CMF = 0.81), and run-off-road crashes (CMF = 0.56). Overall, both countermeasures significantly reduce run-off-road crashes and other injury severities, including total injury, fatal, serious, and moderate injury, and fatal and serious injury crashes. W-beam guardrails have been found to provide greater reductions in run-off-road crash occurrences, whereas clear zone improvements are more effective in reducing injury severity by offering more forgiving roadside conditions.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"222 ","pages":"Article 108218"},"PeriodicalIF":6.2,"publicationDate":"2025-09-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144933165","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Assessment of the safety benefits of HUD warning under high-risk pedestrian crossing event in the connected environment","authors":"Yu Zhang ,&nbsp;Xiaohua Zhao ,&nbsp;Yang Bian ,&nbsp;Jianling Huang ,&nbsp;Duan Yu ,&nbsp;Haolin Chen","doi":"10.1016/j.aap.2025.108157","DOIUrl":"10.1016/j.aap.2025.108157","url":null,"abstract":"<div><div>The head-up display (HUD) warning system in a connected environment is expected to improve driving behavior and enhance pedestrian crossing safety. While existing research has preliminarily examined the effectiveness of HUD warning system in avoiding pedestrian collisions, scant attention has been given to the microcosmic influence on driving behavior and a precise quantification of its overall benefits, especially in high-risk pedestrian crossing scenarios. To investigate these influences, this study employed driving simulations to construct six connected scenarios: three warning systems (Baseline/head-down display(HDD)/HUD) × two weather conditions (clear weather/foggy weather). Data on the driving behavior of 34 drivers across these scenarios were collected. The whole spatial change process of driving behavior in pedestrian crossing events is described from the microscopic level, and the influence law of warning system and weather conditions on reaction, acceleration and deceleration behavior is analyzed. A comprehensive index system reflecting the safety level of the risk-avoidance stage, the recovery level of the recovery stage and the stability level of the overall stage was constructed to explore impact characteristics and utilities of the three warning systems under different weather conditions. The study found that the speed space variations under HDD and HUD conditions were more gentle compared to Baseline conditions, especially in the HUD group, but there were differences in individual adherence to HDD and HUD systems. The results of two-way repeated measures ANOVA and fuzzy comprehensive evaluation indicated that compared to the Baseline and HDD, the HUD warning system improves the safety level and stability level under clear and foggy weather conditions, but does not have a significant advantage in the recovery level. Specifically, the HUD system enables drivers to react earlier, complete risk avoidance earlier, execute smoother acceleration and deceleration maneuvers, and maintain more stable lateral control. Overall, the HUD warning system helps drivers achieve optimal driving performance in a connected environment, even in more hazardous foggy conditions. The research results can provide support for relevant departments to evaluate and improve HUD system in a targeted manner.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"222 ","pages":"Article 108157"},"PeriodicalIF":6.2,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144922790","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Temporal heterogeneity in traffic crash delays: causal inference from multi-scale time factors and sample-wise structural decomposition","authors":"Jianyu Wang ,&nbsp;Kun Qie ,&nbsp;Yang Yang ,&nbsp;Zhiyuan Sun ,&nbsp;Wei Zhou ,&nbsp;Xiantian Chen","doi":"10.1016/j.aap.2025.108220","DOIUrl":"10.1016/j.aap.2025.108220","url":null,"abstract":"<div><div>Traffic crashes remain a critical global public safety concern, exhibiting complex heterogeneity across multiple temporal scales that limits the timeliness and precision of current traffic safety management strategies. Modeling the underlying causation mechanisms across temporal scales presents several challenges, including weak sequential patterns and nonlinear interactions among temporal features. To address these issues, this research categorizes crash data into a macro-scale (annual seasonality) and micro-scale (daily peak intervals and daytime/nighttime variation), and proposes a Multi-Channel Feature Correlation Transformer (MCFformer) to systematically model the influence of multi-scale temporal factors on crash-induced delay. The model introduces a Multi-Channel Feature Correlation (MCFC) mechanism to capture inter-scale couplings, a Recurrent Retention Attention (RRA) module to enhance cross-sample nonlinear dependency modeling, and an attention-based causality interpretation approach to derive the dynamic contribution of each factor. Experimental results demonstrate that MCFformer significantly outperforms mainstream models (e.g., XGBoost and CatBoost) in crash delay regression tasks, achieving a 44% improvement in predictive accuracy and over 26% reduction in RMSE under weakly sequential, non-stationary conditions. Further analysis reveals that environmental factors exhibit higher influence during peak intervals and daytime/nighttime periods, with an average contribution of 55.6%, compared to 51.1% for built environment features. Conversely, in the seasonal dimension, built environment factors contribute 49% on average, exceeding that of environmental factors (44.6%). These findings highlight the scale-sensitive and structurally heterogeneous nature of crash causation, and validate the effectiveness of multi-scale temporal modeling. The proposed framework offers both predictive performance and interpretability, providing theoretical insights and practical guidance for the development of dynamic and refined traffic safety interventions.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"222 ","pages":"Article 108220"},"PeriodicalIF":6.2,"publicationDate":"2025-08-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144916587","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Safety space index (SSI): A two-dimensional metric for quantifying drivers’ perceived risk","authors":"Renjing Tang ,&nbsp;Guangquan Lu ,&nbsp;Jinghua Wang ,&nbsp;Pengrui Li ,&nbsp;Mingyue Zhu ,&nbsp;Miaomiao Liu","doi":"10.1016/j.aap.2025.108216","DOIUrl":"10.1016/j.aap.2025.108216","url":null,"abstract":"<div><div>With the advancement of autonomous driving, there is increasing demand for systems that mimic human decision-making in complex traffic environments. Modeling such behavior requires understanding drivers’ cognitive mechanisms during dynamic interactions. Subjective risk quantification is a key link between perception and decision-making, impacting the system’s ability to generate human-aligned responses. However, existing risk quantification methods predominantly emphasize objective risk assessment or are limited to one-dimensional subjective risk quantification, lacking effective metrics that can comprehensively characterize generalized subjective risk perception in two-dimensional scenarios. To address this gap, this study proposes a novel two-dimensional risk perception metric, the Safety Space Index (SSI), which integrates psychological safe space theory and risk field modeling to quantify drivers’ subjective risk levels. Experimental results show SSI improves correlation with car-following behavior by 32.2%, and achieves a reaction time calibration of 0.92 s. Moreover, SSI effectively distinguishes differences in perceived risk among drivers facing the same conflict scenarios, reflecting strong alignment with human cognitive processes. Extended analyses further reveal that SSI captures the risk homeostasis characteristic of driving behavior, exhibiting centrally clustered target levels that follow a normal distribution in typical scenarios. Additionally, SSI demonstrates robust cross-scenario generalization, maintaining an average target level of 0.50, thereby affirming its adaptability and scalability. As a powerful tool for characterizing drivers’ subjective risk perception in two-dimensional dynamic environments, SSI offers critical theoretical support for human-like behavior modeling, autonomous decision-making strategies, and validation frameworks in intelligent driving systems.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"222 ","pages":"Article 108216"},"PeriodicalIF":6.2,"publicationDate":"2025-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144911667","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A novel data-driven multi-agent pedestrian flow risk assessment framework for avoiding stampede incident","authors":"Zi-Xuan Zhou ,&nbsp;Kai Liu ,&nbsp;Pei-Yang Wu ,&nbsp;Wataru Nakanishi ,&nbsp;Yasuo Asakura","doi":"10.1016/j.aap.2025.108207","DOIUrl":"10.1016/j.aap.2025.108207","url":null,"abstract":"<div><div>This paper addresses the critical issue of monitoring high-density crowds in public spaces like transportation hubs to prevent accidents from overcrowding. It highlights the limitations of prevailing simulation tools in dealing with real-world challenges such as diverse pedestrian destinations, multi-directional flows, and the medley space designs in communal areas. The paper aims to introduce a data-driven, multi-agent framework that assesses crowd dynamics and early warning conditions in different spatial layouts. The model utilizes real-time visual information and reinforcement learning for decision-making, employing a self-iterative algorithm for trajectory planning that aligns with real-world movement characteristics. It enhances model compatibility across various scenarios without the need for parameter fine-tuning. The analysis shows the model’s ability to accurately reproduce pedestrian flow motion in diverse scenarios and indicates a discontinuous state transition in pedestrian flow as density increases. A method for detecting building traffic capacity is proposed, which can identify the threshold of stable pedestrian flow that various spatial arrangements can accommodate, thereby allowing for the advance setting of crowding warning levels. The study suggests that rational spatial layout and information guidance can significantly improve spatial mobility and reduce the risk of crowd stampedes, without expanding the area of architectural spaces.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"222 ","pages":"Article 108207"},"PeriodicalIF":6.2,"publicationDate":"2025-08-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144911666","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"How long is long enough for traffic conflict observation: An investigation using extreme value theory approaches","authors":"Lai Zheng ,&nbsp;Jiayi Li ,&nbsp;Wei Wei ,&nbsp;Carmelo D’Agostino ,&nbsp;Aliaksei Laureshyn","doi":"10.1016/j.aap.2025.108215","DOIUrl":"10.1016/j.aap.2025.108215","url":null,"abstract":"<div><div>Determining the appropriate duration for traffic conflict observation remains a critical yet unresolved challenge in road safety analysis. Existing approaches lack a quantitative approach to determine adequate sample sizes for general conflict-based applications. This study addresses this gap by proposing an extreme value theory based framework to determine appropriate observation durations, based on the inherent stability of road entity safety for a specified period. Using approximately 50 hours of high-resolution LiDAR data from a signalized intersection in Harbin, China, conventional generalized Pareto distribution (GPD) models and Bayesian hierarchical GPD (BH_GPD) models were developed, considering variations in conflict types, intersection approaches, and sample sizes. Two safety indicators, the expected annual number of crashes and the crash return level, both derived from the GPD distribution, were employed to assess sample adequacy based on their respective rates of convergence. Results show that the crash return level, unlike traditional crash frequency metrics, remains non-zero and sensitive to observation duration even in low-risk scenarios, facilitating a more robust identification of adequate sample sizes. Notably, the BH_GPD model generally reduced required observation durations compared to standalone GPD models, particularly for low-conflict scenarios (<em>i.e.</em>, less than 10 conflicts/hour), while yielding narrower credible intervals due to its ability to pool data across sites. A key finding reveals that adequate sample sizes range from 15 to over 45 hours for different scenarios, inversely correlated with conflict rates. This study establishes a quantitative framework to determine the adequate sample size of traffic conflicts, which has the potential to contribute significantly to the standardization of traffic conflict techniques in road safety research.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"221 ","pages":"Article 108215"},"PeriodicalIF":6.2,"publicationDate":"2025-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908484","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Day and night performance differences in detection and deceleration by pedestrian automatic emergency braking systems","authors":"Zeinab Bayati,&nbsp;Asad J. Khattak,&nbsp;Nastaran Moradloo","doi":"10.1016/j.aap.2025.108221","DOIUrl":"10.1016/j.aap.2025.108221","url":null,"abstract":"<div><div>The alarming increase in pedestrian fatalities highlights the urgent need for effective protection technologies. One such technology is the Pedestrian Automatic Emergency Braking (P-AEB) system, a driver assistance feature that provides temporary braking support to help prevent crashes. Despite their availability, the effectiveness of the P-AEB system varies significantly, necessitating continuous performance evaluations under various conditions, especially during nighttime, when 75% of fatal pedestrian crashes occur. This study evaluates the effectiveness of P-AEB systems under varying visibility conditions by analyzing 2,494 tests, comprising 1,654 nighttime tests and 840 daytime tests, conducted with 42 vehicles manufactured between 2021 and 2024. The experimental data were sourced from the Insurance Institute for Highway Safety. A random-effects Heckman sample selection model estimates the detection probabilities and deceleration rates, accounting for unobserved heterogeneity across vehicles and test scenarios. The results show that the detection rates were approximately 98% during the day, 87% at night under low-beam headlights, and 93% at night under high-beam headlights. Despite detection, crashes still occurred in 23% of low beam tests, compared to just 10% during the day. Additionally, crashes at night generally occurred at higher speeds. Furthermore, the model’s results show that halogen low beams can reduce detection capability by up to 43% compared to daytime, underscoring the need for improved P-AEB performance under low-light conditions. This study also incorporates pedestrian movement and vehicle characteristics, such as fuel type and size, revealing their notable association with P-AEB performance. The findings aim to improve pedestrian safety by enhancing P-AEB system effectiveness across varying lighting conditions.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"221 ","pages":"Article 108221"},"PeriodicalIF":6.2,"publicationDate":"2025-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144913140","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A copula-based multivariate extreme value framework for roundabout safety evaluation under mixed traffic","authors":"Abhijnan Maji,&nbsp;Indrajit Ghosh","doi":"10.1016/j.aap.2025.108219","DOIUrl":"10.1016/j.aap.2025.108219","url":null,"abstract":"<div><div>Roundabouts in low- and middle-income countries are not as safe as expected due to non-lane-based traffic behaviors and heterogeneity in traffic conditions. To address the limitations of crash-based analyses, this study developed a proactive, data-driven framework that integrates high-resolution drone-recorded video-based trajectory extraction, multivariate Extreme Value Theory (EVT)-Peak-Over-Threshold (POT) modeling, and probabilistic clustering to identify and classify conflict events at unsignalized roundabouts. Trajectories from videos collected at 22 roundabouts were extracted via advanced computer-vision algorithms and processed in the Surrogate Safety Assessment Model (SSAM) developed by the Federal Highway Administration to compute four surrogate safety measures (SSMs): Time-to-Collision (TTC), Post-Encroachment Time (PET), maximum deceleration (MaxD), and maximum post-collision (hypothetical) velocity change (MaxDeltaV). The quadrivariate EVT-POT model with Gumbel-Hougaard copula was developed to capture joint exceedances of the SSMs and determine context-specific thresholds, i.e., 1.5 s for TTC and PET, −3.0 m/s² for MaxD, and 4.5 m/s for MaxDeltaV, via Mean Residual Life, Threshold Stability, and AIC plots. The copula captured tail dependencies among the SSMs efficiently, marked by its goodness-of-fit diagnostics. Conflicts were mapped spatially, revealing that lane-change interactions constituted ∼ 43 %, rear-end ∼ 38 %, and crossing ∼ 19 % of conflicts, with distinct clustering at approach legs, weaving zones, and pedestrian/bicyclists crossing points. Latent profile analysis using the Gaussian Mixture Model stratified conflicts into five severity levels, i.e., from minor (29.7 %) to critical (7.6 %), enabling prioritized intervention strategies. This framework offers a scalable tool for practitioners to pinpoint high-risk areas and deploy targeted safety countermeasures, enhancing proactive roundabout safety under mixed-traffic conditions.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"221 ","pages":"Article 108219"},"PeriodicalIF":6.2,"publicationDate":"2025-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144903535","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Refining the classification of combined alignment sections on mountainous freeways and analyzing the spatio-temporal effects on crash frequency","authors":"Yesihati Azati ,&nbsp;Xuesong Wang ,&nbsp;Xinchen Ye ,&nbsp;Kaili Zhang","doi":"10.1016/j.aap.2025.108222","DOIUrl":"10.1016/j.aap.2025.108222","url":null,"abstract":"<div><div>Combined alignment sections of mountainous freeways often feature complex geometric configurations—such as downhill sag/convex curves, slope-changing curves, and uphill curves—that significantly affect crash risk. Existing studies typically apply homogeneous segmentation and broad classifications (e.g., downhill, uphill, sag/convex), which fail to capture the specific effects of geometric combinations on crash frequency. In addition, traffic operations and weather conditions in mountainous areas exhibit strong seasonal variation, and using annual data may obscure important patterns, making quarterly analysis necessary. The interaction of complex geometry, dynamic traffic flow, and adverse winter weather results in nonlinear spatio-temporal effects that conventional models cannot effectively capture. To address this, the study integrates road geometry, traffic operation, and environmental data into a Zero-Inflated Negative Binomial (ZINB) model enhanced with Gaussian processes, systematically analyzing the nonlinear spatio-temporal effects on crash frequency. Results show that the proposed model outperforms spatial- or temporal-only models in prediction accuracy (RMSE = 0.566) and model fit (LOOIC = 5961.2), with the variance of spatio-temporal interaction effects estimated at 1.35 (95 % BCI: 1.12–1.58), indicating substantial nonlinear influence. Key findings include a 56 % increase in crash frequency on straight downhill sag curves, a 2 % reduction on straight uphill convex curves, an 80.3 % increase for every additional 1,000 vehicles in daily traffic flow, and a 28.8 % decrease in crash frequency for each 1 °C rise in temperature. The study presents a refined classification and modeling framework that significantly improves crash risk identification and prediction for mountainous freeways, offering strong support for traffic safety management.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"221 ","pages":"Article 108222"},"PeriodicalIF":6.2,"publicationDate":"2025-08-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144903534","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}