Freeway work zones are high-risk locations due to the temporary lane closure. This study aims to estimate rear-end crash risk in these zones based on modeling traffic conflicts. A Bayesian hierarchical random-parameters block maximum (BM) approach, which can accommodate both the multilevel structure and heterogeneity, is proposed to model the extremes of rear-end conflicts. The proposed model is estimated by the traffic conflict data collected from six freeway work zones in Guangdong Province, China. The estimation results indicate that the proportion of oversized vehicles, average speed, and average acceleration have heterogeneous effects on crash risk, whereas traffic volume and the number of open lanes exhibit homogeneous effects. Substantial cross-site heterogeneity is captured by the Bayesian hierarchical framework. Furthermore, model comparison demonstrates that the proposed approach performs better than both the traditional BM approach and the Bayesian hierarchical random-effects BM approach. The findings support the proposed approach as an applicable tool for real-time rear-end crash risk estimation in freeway work zones, which can be readily incorporated into the optimization of proactive safety management strategies, such as variable speed limit and ramp metering.
The rapid expansion of on-demand food delivery has brought algorithmic management, the use of artificial intelligence to oversee labor processes, into everyday working life. The same system that maximizes operational efficiency may also place substantial psychosocial burdens on riders. This study examines the daily dynamics linking algorithmic performance pressure, risky riding behaviors, and near-miss incidents among motorcycle delivery workers. Using a daily diary design, we collected longitudinal data from 50 experienced riders in South Korea over seven consecutive days, yielding 317 person-day observations. A multilevel mediation model decomposed the effects into within-person and between-person components while controlling for individual traits, including baseline risky riding tendencies and work experience. Daily algorithmic performance pressure raised the likelihood of near-miss incidents (total effect), but this relationship ran predominantly through an increase in risky riding behaviors (indirect effect). Once the mediator was included, the direct effect of pressure on near-misses was no longer significant, a pattern consistent with an indirect (mediator-dominant) pathway rather than full mediation in the strict sense. Even with riders' baseline risk tendencies held constant, then, algorithmic pressure compromises traffic safety mainly by inducing rule-violating behaviors as compensatory strategies. These results point to algorithmic architecture as a structural antecedent of occupational hazard. Because the safety risk stems from systemic pressure channeled through behavioral pathways rather than from individual negligence, policy should shift from individual-level enforcement toward systemic reform of platform algorithms, including realistic delivery windows and transparent performance metrics.
Distracted driving is a recognized pedestrian safety risk, yet behavioral heterogeneity within distraction periods remains poorly characterized. We analyze 39,271 one-second observations from 720 naturalistic crosswalk traversals (162 drivers, SHRP2 NDS, 2008-2013) using a Hidden Markov Model (HMM) and Gaussian Mixture Model (GMM) pipeline. Five kinematic states form a near-sequential deceleration cascade, with nondriving-related task (NDRT) engagement concentrated in low-demand states. Monte Carlo cross-validation identified two driver profiles distinguished not by distraction frequency, which differed only negligibly once signal-phase exposure was accounted for, but by regulatory responsiveness: the degree to which approach speed responds to traffic control. The Regulated Tasker (27.7 km/h aggregate approach speed) adjusted speed with the regulatory environment; the Unregulated Tasker (43.9 km/h) responded about a third less strongly. A mixed-effects model on all 162 drivers confirmed that drivers differ significantly in regulatory responsiveness (p=0.024) and that task-active periods were slower, not faster, within driver. An event-triggered analysis of 395 NDRT onsets showed that speed change temporally preceded task engagement, consistent with demand-drop task timing; task-active seconds disrupted the orderly deceleration cascade and shifted initial deceleration onset 102.8 m closer to the crosswalk. The infrastructure-resistant subtype (Unregulated Tasker) is the safety-critical finding: its risk appears less amenable to signal-based control, suggesting physical speed management as a candidate countermeasure to be tested.
Secondary crashes, occurring within the spatiotemporal impact area of primary crashes, tend to exacerbate delays and contribute to further injuries. Existing studies typically model the occurrence or spatiotemporal location (i.e., time gap and distance gap from the primary crash) of secondary crashes separately. However, severe data imbalance remains a major challenge, as the limited availability of secondary crash samples restricts the model's ability to learn minority patterns and generalize effectively. Although data augmentation techniques have been increasingly explored to alleviate this issue, existing approaches often struggle to capture the complex dependencies between dynamic (e.g., traffic flow) and static (e.g., road conditions) features. To address these challenges, we propose VarFusiGAN-Transformer, a hybrid framework that integrates generative modeling and predictive learning for joint prediction of secondary crash occurrence and spatiotemporal location. The proposed VarFusiGAN model employs Long Short-Term Memory (LSTM) networks to enhance the generation of multivariate long-sequence data, while incorporating a static data generator and an auxiliary discriminator to better learn the joint distribution of dynamic and static features. The prediction module performs multi-task joint prediction of both the occurrence and spatiotemporal location of secondary crashes. The proposed framework is evaluated using real-world traffic datasets. To comprehensively assess the quality of generated data, multiple criteria are adopted, including statistical distribution consistency, joint distribution similarity, and inter-variable correlation preservation between real and synthetic data. Experimental results demonstrate that the VarFusiGAN-balanced data significantly enhance classification performance (e.g., G-mean, F1-score, AUC-PR) and spatiotemporal prediction accuracy (e.g., MAE, RMSE). Compared with baseline methods, the proposed framework achieves superior performance in both data generation quality and prediction accuracy, providing an effective solution for secondary crash risk assessment and proactive traffic safety management.
Freeway diverging areas require drivers to decelerate, change lanes, and observe traffic simultaneously, resulting in elevated crash risk compared to basic segments. Previous studies have mainly focused on isolated maneuvers which cannot fully capture the integrated behavioral process during freeway diverging. A multivariate pattern perspective is therefore needed to better understand driver behavior in these areas. Using naturalistic driving data collected in Shanghai, this study investigates short-term driving patterns and their influencing factors in freeway diverging areas, alongside the microscopic interactive environments and longitudinal safety disparities among various driving patterns. A Hierarchical Dirichlet Process Hidden Semi-Markov Model (HDP-HSMM) is used to segment multivariate driving sequences, and the resulting segments are classified into driving patterns using a PCA-based K-means clustering approach. Machine-learning models combined with SHapley Additive exPlanations (SHAP) are then applied to quantify the effects of driver characteristics and environmental factors on the duration ratios of different driving patterns. Three driving patterns are identified: high-speed deceleration (Pattern #0), visual-checking slight lane change (Pattern #1), and low-speed deceleration lane change (Pattern #2). Their spatial evolution shows a consistent trend: Pattern #0 dominates near the entrance of the diverging area and Pattern #2 becomes dominant near the taper section. Driving styles, traffic density, weather conditions, and the number of taper lanes have the strongest effects on pattern durations. Pattern #0 and Pattern #2 prevail in free-flow and car-following conditions respectively, and Pattern #1 shows the highest longitudinal conflict risk, whereas the 3.25-km scenario shows the lowest.
Roadside trees represent the most frequently struck fixed object in fatal crashes in the United States and constitute a substantial safety concern, particularly on high-speed rural roadways. Although trees are often retained or planted along roadways for aesthetic and environmental benefits, their presence within the roadside environment can significantly increase the risk of severe injury outcomes. This study examines the factors influencing driver injury severity in vehicle-tree collisions using a five-year dataset (2020-2024) from the state of Alabama. Following data cleaning and validation, 17,555 crash records were analyzed. Descriptive analysis indicated that fatal and serious injury crashes accounted for 15.20% of cases, with 73.14% occurring in rural areas, and 10.50% involving driving under the influence. To better understand the determinants of injury severity, both traditional regression models and advanced machine learning techniques were employed. The results consistently identified seat belt use as the most influential factor, followed by driver condition (impaired versus non-impaired) and driver gender. Additional significant predictors included crash location, weather conditions, and vehicle type. While these factors have been associated with injury severity in broader crash contexts, their prominence in vehicle-tree collisions reinforces the importance of established safety interventions. From a policy perspective, the findings point to restraint use, impaired driving, and the rural high-speed context as the associations most consistently identified, and are discussed alongside existing roadside design guidance on clear-zone treatment and barrier protection, to mitigate the severity of these high-risk crash events.
Despite the significant hazards posed by distracted driver behaviors, robust monitoring of truck driver activities remains constrained by the scarcity of publicly accessible, high-quality datasets. To fill this gap, we introduce TruckAct, the first naturalistic, multi-source, and multimodal dataset dedicated to truck driver activity recognition. Collected from nine professional drivers under real-world trucking operations without any artificial intervention, TruckAct contains approximately 67 h of synchronized onboard videos, vehicle telemetry, and wearable wristband signals, partitioned into 24,000 non-overlapping 10-second clips with nine well-annotated activity classes. Built on TruckAct, we propose TruckAct-Net, a multi-stream, hierarchical framework that consists of three core components: a Frequency-domain Query Gating (FreqQG) module that suppresses vibration-induced noise in in-cabin video sequences, a dual-stream Time-Frequency TransFormer (TF-Former) encoder that jointly models the temporal evolutions and spectral patterns of vehicle dynamics and wristband signals, and a reliability-guided adaptive fusion strategy that adaptively adjusts the contribution of each modality according to its reliability. Extensive experiments demonstrate that TruckAct-Net achieves robust performance and maintains strong stability even when specific modalities are noisy, occluded, or missing. Visual input delivers discriminative cues, while vehicle motion signals and wristband data help resolve ambiguities and stabilize prediction results. The dataset and source codes are available at https://github.com/WangQF1/TruckAct.
Suburban arterial intersections often face safety challenges related to high speeds, mixed road users, and incomplete traffic control. Signalization is a commonly used countermeasure, but existing studies have primarily examined the signalization of stop-controlled intersections, with limited before-after evidence for previously uncontrolled intersections. This study used crash data from 2020 to 2025 to evaluate the safety effects of signalization at 17 previously uncontrolled suburban arterial intersections in Guangdong Province, China, with 100 untreated intersections serving as the reference group. An empirical Bayes (EB) before-after approach was used as the primary evaluation method, and a safety performance function was developed using the reference intersections to estimate the expected crash frequency at the treated sites in the absence of signalization. A propensity score matching (PSM)-based cross-sectional analysis was also conducted for comparison. Sensitivity analyses based on covariate calibration and E-values were conducted to assess potential confounding bias in the total-crash estimate. Crash modification factors (CMFs) were also estimated by crash type and examined in relation to changes in average speed. The results showed that total crashes increased significantly after signalization, with an EB-CMF of 1.884. Rear-end, same-direction sideswipe, and non-motorized-user crashes also increased significantly, whereas fatal and angle crashes showed no statistically significant changes. The PSM-based cross-sectional analysis produced a directionally consistent crash frequency ratio of 1.913, although its confidence interval was wider. The sensitivity analyses showed that including population density in the SPF changed the total-crash EB-CMF only slightly, from 1.884 to 1.881. The E-value of 3.174 indicated that relatively strong unmeasured confounding would be required to shift the point estimate to the null. Average speed decreased significantly after signalization, but the exploratory analysis did not identify stable associations between the evaluated average-speed measures and the total-crash or crash-type-specific CMFs. Overall, signalization produced varying effects across crash types, highlighting the importance of integrating signal control with supporting measures for left-turn control, pedestrian and non-motorized-user facilities, and approach channelization.
Whether Daylight Saving Time (DST) transitions increase traffic crashes remains a contested policy question, with prior studies reporting 5%-7% increases in fatal crashes over the week following spring-forward but lacking a non-DST control group to isolate the clock-shift effect from seasonal confounding. We address this gap using a difference-in-differences (DiD) design with jurisdictions that do not observe DST as controls. An illustrative city-level analysis uses police-reported crash data from seven U.S. cities (2012-2025), including Tempe, Arizona - which does not observe DST - as a control; this analysis is reported in Appendix A. Our primary analysis uses the Fatality Analysis Reporting System (FARS), covering 375,658 fatal crashes across all 50 U.S. states and the District of Columbia from 2012 to 2022, with Arizona and Hawaii as non-DST controls, in a ±14-day window around each transition. In the illustrative city comparison, adding a non-DST city changes the positive uncontrolled estimate to a near-zero coefficient, although failed diagnostics preclude city-level causal inference. In the primary FARS DiD, for which pre-trend tests do not reject the null of no differential trends, the estimate for overall fatal-crash counts is also null (＋0.2%, p=0.954, ±14-day window). However, morning (6-10 AM) fatal crashes increase by ＋25.9% (p=0.007) in the 6:00-9:59 AM slot averaged over the ±14-day spring-forward window (post-transition days 0-14 vs. pre-transition days -14 to -1)-an effect that strengthens to ＋26.8% (p=0.002) once month-of-year fixed effects and day-of-week-by-DST interactions are added, remains similar when scaled by annual state-level vehicle-miles-traveled (＋22.8%, p=0.016), and is robust to bandwidth as narrow as ±3 days (＋45.8%, p=0.003) and to the exclusion of either non-DST control state. All FARS-based estimates pertain to fatal-crash counts, not crash risk per trip; annual VMT scaling cannot absorb short-run or intraday changes in traffic exposure. The post-transition pattern is not readily attributable to acute sleep deprivation alone and instead points to potentially interacting sleep, circadian, ambient-light, and traffic-exposure channels that this design cannot separately identify. The fall-back effect on total fatal-crash counts likewise vanishes with controls, a pattern consistent with seasonal confounding rather than a clock-shift effect. These findings suggest that prior uncontrolled comparisons may overstate DST's overall crash effect, while underscoring a targeted morning fatal-crash-count increase that merits policy attention.
This study examined differences in the frequency, severity, and distribution of officially recorded traffic sanctions before and after attendance at attitude-change-based road safety awareness and re-education courses integrated into the Penalty Point System in Catalonia. A retrospective longitudinal pre-post design was applied to administrative records from the complete 2017 cohort of 14,022 course participants. The main sanction-profile analyses focused on 9,886 drivers with at least one pre-course sanction, accounting for 36,546 pre-course and 23,022 post-course sanctions. Period-by-category differences were examined using Pearson chi-square tests of homogeneity, Cramér's V, and aggregate log-linear Poisson models. During follow-up, 37.1% of drivers with previous sanctions had no further officially recorded sanction, and the total sanction count in this group was 37.0% lower. Alcohol- and drug-related sanctions showed the largest relative decline (-77.9%), followed by safety-device and driver-related sanctions. Mobile phone-related sanctions declined by 62.0%. Speeding showed a smaller reduction (-17.2%) and represented a larger share of the post-course profile. Significant distributional differences were found for the main offence categories, offence severity, driver-related offence subtypes, and serious speeding subtypes (Cramér's V = 0.128-.210). Sanctions recorded in average-speed enforcement sections and urban access areas increased despite declines in other speeding categories. These findings show that a single recidivism indicator would conceal meaningful differences in offence type, severity, and speeding context. Because the study was observational, lacked a comparison group, and relied on detected offences, the patterns cannot be attributed causally to course attendance or interpreted as direct evidence of behavioural or attitudinal change.
In complex traffic systems, driving risk often evolves in a continuous and progressive manner prior to crash occurrence. How to effectively represent and analyze such latent risk states remains a central challenge in traffic safety research. In recent years, risk field-based approaches have introduced spatial and spatiotemporal continuous modeling paradigms, providing new perspectives for characterizing the distribution of traffic risk and its dynamic evolution. Motivated by the rapid growth of this research area and the lack of a systematic synthesis, this paper presents a comprehensive review of studies applying risk field theory to driving safety and traffic risk analysis. Following the PRISMA guidelines, relevant literature was collected through multi-database searches and analyzed using a combination of bibliometric analysis and qualitative review. The review systematically summarizes the theoretical foundations, modeling elements, data sources, analytical methods, and application domains of risk field-related research. Particular attention is given to studies that conceptualize traffic risk as a continuous field, complemented by a broader review of traffic risk factor literature to identify key elements and analytical dimensions involved in risk field modeling. On this basis, the paper synthesizes research progress in major application areas, including traffic safety state representation, driving behavior analysis, traffic conflict assessment, and autonomous driving and human-machine cooperative systems. Differences and commonalities among existing studies are compared in terms of modeling strategies, data support, and application scenarios. Through this systematic review, the paper clarifies the main research themes and methodological trends of risk field-based studies, providing a structured framework for understanding the evolution and application of this approach and offering methodological insights for risk perception modeling and safety-oriented decision support in intelligent transportation systems (ITS).
Driving risk assessment is crucial for advancing transportation safety, particularly in the context of increasing vehicle automation and electrification. This study proposes a modified two-dimensional time-to-collision (2D-MTTC) measure that incorporates relative acceleration in two-dimensional space to better characterize driving risk across different types of passenger vehicles. The proposed 2D-MTTC is validated through a two-stage validation process using different datasets to show its effectiveness and performance compared with the benchmark two-dimensional time-to-collision (2D-TTC): i) macro-level validation by computing the correlation between conflict risk and historical crash records of the road network; and ii) micro-level validation by detecting critical risk events in real conflicts and crash trajectories. After validation, the driving risk captured by 2D-MTTC between different types of passenger vehicles is examined using the Third Generation Simulation dataset (TGSIM). Specifically, the identified driving risks are compared between automated vehicles (AVs) and human-driven vehicles (HDVs), as well as between electric-based automated vehicles (EAVs) and internal combustion engine-based automated vehicles (ICEAVs). The results demonstrate the effectiveness of 2D-MTTC and its improved performance relative to 2D-TTC in identifying driving risks and capturing their duration and intensity. The findings further reveal that the driving risk characteristics among AVs and HDVs, EAVs and ICEAVs are significantly different and urge the consideration of vehicle types in proactive safety analysis.
This study develops a cross-facility hierarchical Bayesian marginal-copula framework for jointly modeling rear-end crash counts and three connected-vehicle (CV)-derived longitudinal conflict measures: time to collision (TTC) conflicts, deceleration rate to avoid a crash (DRAC) conflicts, and time to collision with disturbance (TTCD)-based conflict risk. The framework is designed for cross-facility generalization between freeways and arterials: hierarchical marginal models with facility-type random effects enable partial pooling and principled information sharing across facility types, while an inference-for-margins (IFM)-estimated multivariate Gaussian copula characterizes crash-conflict associations among the four safety indicators while preserving outcome-specific marginal distributions. The empirical analysis combines police-reported rear-end crashes with CV trajectory data aggregated to freeway and arterial segments in Ann Arbor, Michigan. Results show strong exposure effects: an e-fold increase in annual average daily traffic (AADT) is associated with an approximately threefold increase in expected crash counts, whereas an e-fold increase in CV exposure increases TTC and DRAC conflict counts by about 2.5 to 2.6 times and TTCD-based conflict risk by roughly a factor of 4 to 5. The conflict measures are more strongly concordant with one another (Kendall's τ≈0.47 for TTC-DRAC) than with crashes (all crash-conflict τ<0.20), suggesting that TTC and DRAC capture closely related longitudinal interaction mechanisms whereas crash occurrence is only weakly associated with any single conflict metric. Cross-facility prediction experiments show that naive freeway-to-arterial extrapolation can substantially degrade predictive performance; for example, TTC-conflict root mean square error (RMSE) increases by about 89%. By contrast, the hierarchical pooled specification mitigates this degradation and achieves arterial predictive accuracy comparable to an arterial-only benchmark, with sensitivity results indicating stable performance under moderately reduced arterial training samples.
Existing approaches for concrete scenario generation and assessment primarily focus on trajectory-level interactions while neglecting physical perception constraints, such as sensor field-of-view limitations and occlusions from surrounding objects, leading to unrealistic risk representations. To fill this gap, this study proposes a framework that integrates vehicle dynamics, environmental constraints, and perception limitations into a unified generation process. Specifically, vehicle trajectories are parameterized using a kinematic bicycle model and perturbed through controlled variations in longitudinal acceleration and steering angle. A voxel-based surrogate perception metric is introduced to efficiently evaluate cooperative perception (CoP) by modeling line-of-sight occlusions, which achieves a balance between computational efficiency and accuracy. By modeling perception limitations, the framework generates not only risky but also CoP-adverse scenarios. A multi-constraint optimization framework driven by the genetic algorithm is proposed, which ensures that generated scenarios satisfy physical feasibility, road compliance, collision avoidance, and degraded perception conditions, while increasing interaction risk between agents. Experimental results based on real-world trajectory datasets demonstrate that the proposed approach can effectively generate diverse and high-risk scenarios with reduced CoP performance and satisfactory computational efficiency. This framework may not only help generate CoP-adverse scenarios for the assessment and evaluation of vehicle-infrastructure CoP systems, but also offer a tool for the construction of scenario library to support end-to-end learning-based methods.
In the context of connected vehicles, head-up displays (HUD) have been widely used to provide warning information and enhance driving safety. However, existing systems still lack effective ways to present information about inherent road risk, which is particularly crucial for truck drivers in complex road scenarios. To address insufficient support for truck drivers' evasive maneuvers in scenarios involving overlapping road-vehicle risk factors, this study focused on a high-risk situation involving sudden braking by a preceding vehicle on a curve. Based on the International Road Assessment Programme (iRAP) methodology, a Road-risk Head-Up Display (RHUD) warning system integrating quantified road-risk information was developed. A driving simulation experiment was conducted using a scenario based on the Qingyin Expressway prototype, in which a traditional HUD and an enhanced RHUD were implemented for comparison. Thirty-seven professional truck drivers completed simulated driving tasks involving sudden lead-vehicle braking on a curved segment under both display conditions. The longitudinal safety margin index was selected as the primary metric for risk-avoidance behavior. Survival analysis showed that, in the scenario of sudden braking before a curve, the RHUD significantly improved drivers' risk-avoidance behavior (χ2 = 7, p = 0.008), yielding improvements of 27.68%, 32.67%, and 37.94% at the 25th, 50th, and 75th percentiles, respectively. Further analysis revealed that under the RHUD condition, the effects of demographic factors such as age were substantially attenuated, whereas driving performance was primarily governed by situational variables, particularly driving experience and initial speed. This suggests that structured road-risk information helps reduce individual differences in drivers' responses to high-risk events. These findings demonstrate the effectiveness of incorporating road-risk information into truck HUD warning design and confirm the feasibility of linking iRAP-based quantitative road-risk information with in-vehicle warning systems. The proposed approach provides empirical support for optimizing truck warning strategies based on road-risk assessment results and offers practical implications for improving safety in complex road environments.
Multimodal human-machine interaction (HMI) is increasingly implemented in smart cockpit systems, yet empirical evidence on how specific modality configurations jointly shape driver workload and performance remains limited. From a neuroergonomic perspective, this study examined how variations in combinations of visual layout, auditory feedback, and haptic interaction were associated with driver workload regulation and vehicle-control performance in a controlled driving simulation. Twenty-nine licensed drivers performed secondary interaction tasks under systematically varied multimodal configurations, while neural activity, eye movements, physiological responses, driving performance, and subjective workload were concurrently recorded. The results indicate that different multimodal configurations do not produce uniform patterns of workload and performance outcomes; instead, different configurations elicit distinct patterns across neural, physiological, and behavioral indicators, suggesting differences in how cognitive demands may be distributed across perceptual, motor, and executive processes. Configurations characterized by more stable visual attention and direct interaction were associated with different patterns of processing demands and vehicle-control responses. Cross-indicator analysis further showed that relationships among measurement domains were selective rather than uniformly convergent. Notably, subjective workload did not always align with objective indicators, suggesting that different measures may capture partially distinct aspects of driver state. Overall, the findings highlight that the effectiveness of multimodal HMI depends on how modalities are configured rather than simply on the presence of multiple modalities, and underscore the value of interpreting driver workload through complementary rather than interchangeable indicators.
Analyzing the factors influencing pedestrian crashes at crosswalks from both drivers' and pedestrians' visual perspectives is essential for improving pedestrian safety. However, current studies have not sufficiently considered the holistic visual characteristics of road environments and have provided limited practical guidance for crosswalk safety improvement. To address these gaps, this study integrates XGBoost with SHAP values to conduct an interpretable analysis of pedestrian crashes at crosswalks, considering holistic characteristics of the visual road environment. Appearance, depth, and color features were extracted from street-view images for 373 crosswalks, and hierarchical clustering was applied to classify them into three distinct visual clusters representing holistic environmental characteristics. These visual clusters were then combined with road design, crosswalk, traffic control, exposure, and contextual variables to develop an XGBoost model for predicting whether pedestrian crashes occurred at each crosswalk, achieving an accuracy of 0.866 and an AUC of 0.896. SHAP values were then used to identify the relative importance of each independent variable, the specific effects of individual variables, and the joint effects between variables. Road width and visual clusters emerged as important factors, with wider roads associated with a higher predicted crash likelihood and visually balanced environments associated with a lower predicted crash likelihood. Notable joint effects were also observed between road width and visual clusters. The proposed methodology helps provide practical insights for crosswalk design and pedestrian safety improvement.
Safe operation of Autonomous Vehicles (AVs) in complex traffic scenarios remains a challenge, particularly in interactions with Powered Two-Wheelers (PTWs), whose riders represent a common and vulnerable group of road users. This study proposes DRIVE (Diffusion-Reachability-based Interaction and Validation for Efficient scenario testing), an integrated framework for risk-oriented accelerated safety testing in car-PTW interactions. The framework combines diffusion-based trajectory generation, backward reachability analysis, and reachability-guided sampling to construct and prioritize safety-critical scenarios. Based on 314 reconstructed real-world crashes from the in-depth crash database, a diffusion model produced 6280 realistic interaction trajectories. The generated trajectories were represented using a truncated Gaussian mixture model and organized into five risk levels through reachability analysis. A reachability-guided sampling scheme then allocated simulation effort to dynamically critical regions. Simulation tests with a production-level automated driving system show that DRIVE yields crash rates of 84.00%-94.47% in the targeted risk sets, compared with 15.29% for replayed crashes and 12.53%-24.80% for baseline generative methods. At the same time, DRIVE covers a wider range of severe crashes. These results show that DRIVE improves failure discovery and targeted stress-testing efficiency under a fixed simulation budget, so that comparable safety conclusions can be drawn from far fewer simulations in PTW interaction scenarios.
