Accident; analysis and prevention最新文献

{"title":"GeoShapley-based interpretation of older adult pedestrian fatal vs injury crash frequency in dense urban environments","authors":"I Gede Brawiswa Putra ,&nbsp;Pei-Fen Kuo ,&nbsp;Febrian Fitryanik Susanta ,&nbsp;Bimo Harya Tedjo ,&nbsp;Dominique Lord","doi":"10.1016/j.aap.2026.108450","DOIUrl":"10.1016/j.aap.2026.108450","url":null,"abstract":"<div><div>As the world’s population ages, ensuring the safety of older adult pedestrians has become an urgent priority in transportation planning. However, most existing studies rely on global models that overlook spatial heterogeneity and fail to capture nonlinear, location-specific interactions between the environmental factors and crash outcomes. Moreover, subjective perceptions (e.g., how safe or walkable an area feels) may influence pedestrian behavior and crash exposure but are underexplored in traffic safety research.</div><div>This study addresses these gaps by integrating subjective perception indicators extracted from Street View Images (SVI) with machine learning models to examine the severity of older adult pedestrian crashes at intersections in Taipei City. Three modeling frameworks are evaluated and compared: global Negative Binomial Regression (NBR), Geographically Weighted Negative Binomial Regression (GWNBR), and GeoShapley, a spatially interpretable extension of the SHAP framework for XGBoost. A total of 36 environmental and perceptual variables are evaluated in relation to injury and fatal crash frequencies.</div><div>Among these models, GeoShapley achieved the best performance and revealed that spatial location (GEO) and its interactions with environmental factors and subjective perceptions were among the most influential predictors. In some areas, higher walkability was associated with reduced injury crash frequencies, especially in older and urban districts. In addition, the effect of convenience stores and nursing homes on the frequency of fatal crashes varied significantly across locations, reflecting the spatial clustering of pedestrian activity and older adults. Overall, the findings demonstrate the value of spatially explicit machine learning tools and subjective perceptions in understanding localized crash dynamics in aging urban populations.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"230 ","pages":"Article 108450"},"PeriodicalIF":6.2,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146177362","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":"Study on a multi-factor lane-changing risk resilience assessment model based on genetic algorithm and fault tree analysis","authors":"Qiang Luo ,&nbsp;Haihui Wang ,&nbsp;Junheng Yang ,&nbsp;Xiaodong Zang ,&nbsp;Xinqiang Chen ,&nbsp;Octavian Postolache","doi":"10.1016/j.aap.2026.108443","DOIUrl":"10.1016/j.aap.2026.108443","url":null,"abstract":"<div><div>Current lane-change risk assessment models often lack dynamic adaptation to adverse weather and validation against real-world outcomes. To bridge this gap, this study re-frames the problem through a resilience engineering lens, defining risk resilience as the lane-changing system’s capacity to absorb weather disturbances and maintain safety through adaptation. To operationalize this concept, we introduce two complementary metrics: the Risk Exposure Level (REL) and the Risk Severity Level (RSL). We propose a weather-aware, resilience-oriented assessment framework that integrates a Genetic Algorithm (GA)-calibrated Stopping Sight Distance (SSD) model with Fault Tree Analysis (FTA). Using the CitySim naturalistic driving dataset, a dual-threshold identification algorithm was applied to extract 310 lane-change events (218 sunny, 92 rainy). Key influencing factors, including weather, surrounding vehicle distribution, lane-change direction, and location, were identified through statistical testing. The GA was employed to optimize critical braking parameters (deceleration, reaction time) in the SSD/SDI model, enabling self-adaptive risk thresholds under different weather conditions. REL and RSL quantify the probability (exposure) and severity (consequence) of conflicts from multiple vehicle groups, which are systematically integrated via FTA to assess overall system robustness. Model calibration and testing using trajectory data showed a 42.38% improvement in fitness over the baseline model. A PyQt5-based visualization platform was developed to support practical application. The results confirm that the model effectively captures real-time lane-changing risk, providing a reliable tool for proactive safety management and resilience-oriented decision support in intelligent transportation systems.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"230 ","pages":"Article 108443"},"PeriodicalIF":6.2,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146187583","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":"Enhancing legal driving for autonomous vehicles through law-compliance potential fields","authors":"Chengxiang Zhao ,&nbsp;Zichun Xiong ,&nbsp;Lei Song ,&nbsp;Xiaohan Ma ,&nbsp;Zhengxian Chen ,&nbsp;Siyu Wu ,&nbsp;Jun Li ,&nbsp;Wenhao Yu ,&nbsp;Hong Wang","doi":"10.1016/j.aap.2026.108441","DOIUrl":"10.1016/j.aap.2026.108441","url":null,"abstract":"<div><div>With the advent of the mixed traffic era, ensuring autonomous vehicles’ compliance with traffic laws alongside human drivers has become increasingly critical. Existing decision-making methods predominantly emphasize safety, inadequately addressing systematic compliance with traffic laws, leading to potential legal violations in complex driving scenarios. To bridge this gap, this paper proposes a comprehensive Law-Compliance Potential Fields-based method. Traffic law constraints are systematically categorized into four potential fields, which explicitly encode vehicle states, static and dynamic elements, and compliance thresholds. A novel fusion strategy is further designed to effectively resolve field-overlap distortions. Finally, the constructed law-compliance potential fields are integrated into a model predictive control-based decision-making framework, and five representative scenarios are designed for experimental validation, including a critical scenario of safety-compliance conflict. The evaluated results of scenario tests demonstrate that the proposed method markedly enhances autonomous vehicles’ compliance capabilities, effectively balancing safety considerations even under challenging driving conditions.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"230 ","pages":"Article 108441"},"PeriodicalIF":6.2,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146187588","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":"CoDEA: A framework for extraction and augmentation of cooperative lane-changing scenarios from naturalistic driving data","authors":"Ye Li ,&nbsp;Weiran Li ,&nbsp;Rui Zhou ,&nbsp;Lu Xing ,&nbsp;Changyin Dong ,&nbsp;Dan Wu","doi":"10.1016/j.aap.2026.108451","DOIUrl":"10.1016/j.aap.2026.108451","url":null,"abstract":"<div><div>As modern transportation systems face increasing complexity, with challenges such as increased vehicle volumes, limited road resources, and rising safety concerns, there is an urgent need for innovative solutions. Cooperative driving, which enables vehicles to share information and collaborate through communication technologies, presents a promising solution to enhance safety, reduce congestion, and improve mobility. However, the validation of cooperative driving systems is hindered by a critical scarcity of real-world data. To address this challenge, we introduce CoDEA (Cooperative Driving Extraction and Augmentation), a comprehensive three-stage pipeline designed to generate robust and realistic cooperative driving datasets. First, a systematic method is developed to extract cooperative lane-changing behaviors from large-scale Naturalistic Driving Data (NDD), ensuring that the extracted data captures the key kinematic and cooperative features of real-world scenarios. Next, to effectively generate realistic cooperative lane-changing scenarios, we enhance the DiffTraj framework by introducing our Interaction-Aware Context Encoding (IA-CE) module. This module allows the diffusion model to condition its generation process on the nuanced interactions between vehicles, leading to the creation of more realistic and diverse cooperative trajectories. Finally, the effectiveness of the generated trajectories is evaluated using computational metrics such as RMSE and MAE, and by comparing key feature distributions between real and generated trajectories. The results show a strong similarity between the generated data and real-world cooperative lane-changing patterns, while also introducing greater diversity in certain features. Ultimately, the proposed CoDEA approach lays a solid foundation for advancing cooperative lane change control algorithms by providing a robust dataset for both training and evaluation, effectively bridging the gap between real-world complexity and algorithm testing environments.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"230 ","pages":"Article 108451"},"PeriodicalIF":6.2,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146163419","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":"Modeling crash frequencies at highway-railroad grade crossings in Kentucky in the United States","authors":"Arunabha Banerjee,&nbsp;Kirolos Haleem","doi":"10.1016/j.aap.2026.108452","DOIUrl":"10.1016/j.aap.2026.108452","url":null,"abstract":"<div><div>Previous safety studies at highway-railroad grade crossings (HRGCs) have typically examined crash severities. However, there remains a notable gap in studies that developed safety performance functions (SPFs) (or crash frequency models) at these critical locations. This study develops SPFs for both total and fatal-and-injury (FI) crashes at HRGCs in the state of Kentucky in the U.S. using ten years (2014–2023) of crashes. The study merged extensive crash records and roadway attributes from the Kentucky Transportation Cabinet (KYTC) and Federal Railway Administration (FRA). Additional effort was made to manually collect geometric and infrastructure features nearby HRGCs (e.g., skewness, presence of exclusive turns, presence of channelizing island, and different sign types). The negative binomial “NB”, heterogeneous NB “HTNB”, Conway-Maxwell-Poisson “CMP”, and heterogeneous CMP “HTCMP” models were fitted to handle data over-dispersion. The HTCMP model (with varying dispersion parameter) demonstrated the best performance. The model results showed that urban locations with poor illumination, fewer number of warning bells (≤2), the absence of track signals, and skewed geometry were associated with increased crash frequencies. Angle and rear-end HRGC-related crashes were predominant at the high-crash-prone HRGC locations, often involving left-turn movements and driver distractions. Key risk factors varied between total and FI crashes, with features like the presence of stop lines, the presence of parking structures, skewness, and the number of through lanes differently influencing each model. Based on the model findings and high-crash location analysis, several countermeasures were recommended near HRGCs, e.g., installing high-intensity LED lighting to improve nighttime visibility, and installing static signs (skewed-intersection warning and distracted driving).</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"230 ","pages":"Article 108452"},"PeriodicalIF":6.2,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146177397","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":"Railway drivers’ physiological responses to typical hazardous scenarios: differences between professional drivers and student drivers","authors":"Zizheng Guo ,&nbsp;Junjie Tang ,&nbsp;Mingrui Li ,&nbsp;Jun Zhang ,&nbsp;Yan Zhang ,&nbsp;Guofa Li","doi":"10.1016/j.aap.2026.108455","DOIUrl":"10.1016/j.aap.2026.108455","url":null,"abstract":"<div><div>Unexpected object intrusions on railways present significant safety hazards that can lead to accidents, injuries, and operational disruptions. Railway drivers serve as the critical final line of defense in accident prevention, it is very necessary to study railway drivers’ physiological responses to unexpected object intrusions. Existing studies primarily addresses behavioral responses, with few considering drivers’ physiological cognition responses. This study recruited both professional railway drivers and student participants to investigate drivers’ physiological responses to different hazardous scenarios. Electroencephalogram (EEG) data were collected during driving tasks across four typical hazardous scenarios for time-domain and frequency-domain analyses. Results revealed that professional drivers exhibited greater efficiency in neural resource allocation, cognitive resource integration, executive control and decision-making, as well as visual processing compared to student. Subjective hazard ratings were higher for drivers than for students, indicating greater perceived hazard. Professional drivers displayed distinct response patterns across different hazard scenarios: large-volume hazardous obstacles trigger sustained high cognitive load and executive control activation during the middle and later stages of hazard encounters, with moderate increases observed in the later stage, whereas small-volume hazardous obstacles elicit elevated cognition that remains stable upon hazard detection. Dynamic hazard scenarios elicited stronger visuospatial activation in drivers. Additionally, higher speeds imposed greater cognitive demands on drivers, with enhanced activation of brain regions associated with executive function, control, and decision-making observed during the early stage of hazard encounters. This study advances understanding of expertise-driven neurophysiological responses and provides evidences for developing targeted training programs and neurocognitive frameworks for railway safety enhancement.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"230 ","pages":"Article 108455"},"PeriodicalIF":6.2,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146187587","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":"Trajectory planning for traffic safety with dynamic ethical risk adjustment","authors":"Chengcan Liu ,&nbsp;Haohan Hu ,&nbsp;Tie Ma ,&nbsp;Yuhao Zhang ,&nbsp;Mo Jia ,&nbsp;Linheng Li ,&nbsp;Jing Gan ,&nbsp;Xu Qu ,&nbsp;Bin Ran","doi":"10.1016/j.aap.2026.108425","DOIUrl":"10.1016/j.aap.2026.108425","url":null,"abstract":"<div><div>The integration of ethical principles into the trajectory planning of connected and automated vehicles remains a critical challenge, balancing technical efficacy with societal values. Current algorithms prioritize ego-vehicle safety but inadequately address ethical risks for all road users and cultural variations in moral preferences. This study proposes a trajectory planning algorithm with dynamically adjustable ethical risks, introducing two key innovations: (1) an “ethical knob” mechanism that flexibly weights risks between ego vehicles and other road user, enabling region-specific ethical customization, and (2) a hybrid subjective–objective weighting method combining Analytic Hierarchy Process and coefficient of variation to dynamically allocate weights among four ethical principles—utilitarianism, justice theory, deontology, and responsibility ethics. The algorithm embeds these frameworks into a safety potential field model, translating abstract ethics into computable risk costs. Multi-scenario simulations demonstrate that neutral ethical knob settings minimize overall risk costs, while dynamic weighting automatically adapts to environmental changes compared to static approaches. Crucially, the framework avoids predefined “trolley problem” dilemmas by focusing on accident prevention rather than post-collision decisions. By enabling cultural adaptability and transparent ethical trade-offs, this work advances interdisciplinary solutions for socially acceptable autonomous systems.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"229 ","pages":"Article 108425"},"PeriodicalIF":6.2,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146049043","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":"Visual attention and driving behavior of male autistic individuals while encountering driving hazards: A driving simulator study","authors":"Wondwesen Girma Mamo ,&nbsp;Wael K.M. Alhajyaseen ,&nbsp;Hélène Dirix ,&nbsp;Kris Brijs ,&nbsp;Giovanni Vanroelen ,&nbsp;Qinaat Hussain ,&nbsp;Geert Wets ,&nbsp;Veerle Ross","doi":"10.1016/j.aap.2026.108420","DOIUrl":"10.1016/j.aap.2026.108420","url":null,"abstract":"<div><div>Hazard perception is an important aspect of driving competence that significantly contributes to road safety. Allocating sufficient visual attention to hazards and responding accordingly can help reduce the likelihood of road crashes. Although hazard perception has been investigated to some extent in autistic individuals, little attention is given to hazards for which attention has to be divided among different hazard sources. The current study assessed visual attention and driving behavior of autistic individuals to hazards, including dividing and focusing attention (DF), environmental prediction (EP), and behavioral prediction (BP) hazards. A total of 53, male participants, 19 autistic and 34 non-autistic individuals participated in the study. All participants completed a driving simulator scenario while wearing an eye-tracking system. The included eye-tracking measures were time to first fixation (TTFF), frequency count (FC), first fixation duration (FFD), and average fixation duration (AFD). The included driving measures were brake reaction time (BRT), minimum time-to-collision (minTTC), and speed change immediately before encountering the hazard. A self-reported appraisal regarding difficulty in managing hazards was also included. A series of Linear Mixed Models (LMM) were computed to assess the effects of participant group (autistic and non-autistic) and hazard types (DF, EP and BP) on the included measures. Comparisons of visual attention between autistic and non-autistic participants when responding to hazards yielded mixed results. For certain hazards, autistic participants demonstrated faster fixation (e.g., DF and BP). In contrast, for other hazards, non-autistic participants exhibited quicker fixation (e.g., EP) and longer average fixation duration (e.g., DF and EP). For some hazards, however, both groups displayed comparable levels of average fixation duration (e.g., BP). Although variations in visual attention to hazards were observed between autistic and non-autistic individuals, these differences did not manifest in driving performance metrics. This is evidenced by the absence of significant interactions between participant groups and hazard types concerning driving measures. However, autistic individuals were more likely to experience crashes involving BP hazards than non-autistic individuals. Notably, inexperienced autistic participants had a higher crash rate on BP hazards compared to non-licensed non-autistic participants. In contrast, the crash rates were comparable between licensed participants in both groups. The study may reflect that pre-driver autistic participants could benefit from hazard perception training, particularly in dealing with BP hazards.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"229 ","pages":"Article 108420"},"PeriodicalIF":6.2,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146076717","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":"When feeling safe becomes risky: A VR-EEG-computer vision framework for analyzing cyclist safety in dynamic traffic environment","authors":"Lurong Xu ,&nbsp;Tengfeng Lin ,&nbsp;Steve O’Hern ,&nbsp;Alexa Delbosc ,&nbsp;Zhuo Chen ,&nbsp;Inhi Kim ,&nbsp;Shuli Luo","doi":"10.1016/j.aap.2026.108418","DOIUrl":"10.1016/j.aap.2026.108418","url":null,"abstract":"<div><div>The mismatch between cyclists’ perceived safety and actual crash risk in mixed-traffic environments is a critical yet underexplored issue in road safety research. While prior studies have focused on static environmental factors, they often overlook the real-time influence of dynamic visual stimuli on risk perception. To address this gap, this study developed a multisource-integrated virtual reality (VR) experimental platform that synchronously captured millisecond-level electroencephalography (EEG) signals from 72 participants, built environment (BE) features, and time-to-collision (TTC) data from VISSIM microsimulation. A Long Short-Term Memory (LSTM) model was used to examine how mismatches emerge between perceived safety and crash risk. Results reveal a ‘perceptual relief period’ after being overtaken, where cyclists exhibit higher perceived safety despite persistent threats from following vehicles, creating a potentially hazardous temporal window. This mismatch effect is further amplified in environments characterized by high spatial enclosure, complex visual textures, dense vegetation, and low visible vehicle density. These findings suggest that BE features intended to enhance aesthetic appeal or reduce stress may inadvertently impair cyclists’ ability to perceive risk in high-conflict areas. This study offers empirical support for an integrated human–vehicle–environment safety framework and calls for interdisciplinary collaboration between neuroscience and transport engineering in the design of safer mobility systems.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"229 ","pages":"Article 108418"},"PeriodicalIF":6.2,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146076718","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 graph-based spatio-temporal framework for predicting safety-critical pedestrian–vehicle interactions at unsignalized crosswalks","authors":"Kaliprasana Muduli ,&nbsp;Indrajit Ghosh ,&nbsp;Satish V. Ukkusuri","doi":"10.1016/j.aap.2026.108409","DOIUrl":"10.1016/j.aap.2026.108409","url":null,"abstract":"<div><div>Pedestrian safety remains a critical global concern, especially in countries like India, where unsignalized crossings with limited traffic control contribute to high pedestrian fatality rates. This study proposes a novel graph-based framework for analyzing pedestrian–vehicle interactions, advancing beyond traditional indicator-based, trajectory prediction, and pairwise modeling approaches. Indicator-based methods (e.g., PET, TTC) are retrospective and fail to capture evolving dynamics. Trajectory prediction models suffer from error accumulation, while pairwise models like LSTM or Transformer architectures are limited to two-agent interactions, restricting scalability and scene comprehension. In contrast, the proposed framework constructs holistic, scene-level multi-relational graphs, representing pedestrians and vehicles as interconnected nodes, with explicit modeling of pedestrian–pedestrian, vehicle–vehicle, and pedestrian–vehicle interactions. Unlike existing approaches that treat each pedestrian–vehicle pair in isolation, the proposed method represents all road users present in the scene as nodes in a dynamic spatio-temporal graph, enabling the model to learn not only direct pedestrian–vehicle relationships but also indirect influences mediated by surrounding pedestrians and vehicles. This design eliminates the need to pre-select interaction pairs, simplifying real-world deployment and improving scalability in dense traffic scenarios. Disentangled Multi-Scale Aggregation (DMSA) captures group behavior by focusing on contextually relevant agents, while a temporal CNN backbone models both short- and long-range dependencies efficiently. Empirical evaluations demonstrate the superior performance of the proposed model, which achieved a test accuracy of 90.6%, F1-score of 0.906, precision of 0.927, recall of 0.886, specificity of 0.928, and an AUC of 0.950, outperforming widely used baselines from the literature, such as GRU, LSTM, and Transformer-MLP, that have been applied in pedestrian interaction modeling tasks. Ablation studies confirmed the importance of Multi-Relational Adjacency Matrices (MRAM) and DMSA in improving accuracy and reducing false positives. By modeling scene-level dynamics, the framework enables context-aware prediction of critical events, supporting proactive conflict warning systems.</div></div>","PeriodicalId":6926,"journal":{"name":"Accident; analysis and prevention","volume":"228 ","pages":"Article 108409"},"PeriodicalIF":6.2,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145974327","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}