Accident; analysis and prevention最新文献

As automated vehicles (AVs) become increasingly prevalent in mixed-traffic environments, it is essential to understand how they interact with human-driven vehicles (HDVs), especially in safety-critical situations. Existing research has primarily focused on AVs' collision avoidance strategies, often neglecting how AV maneuvers simultaneously influence the decision-making behaviors of HDVs. This study develops the multi-agent state-space attention-enhanced deep deterministic policy gradient (MA-ASS-DDPG) framework, leveraging the Third Generation Simulation (TGSIM) dataset for the first time to learn interactive car-following behaviors of an AV and the following human-driven vehicles (FHDV) in safety-critical scenarios. By integrating the attention mechanism to dynamically prioritize critical motion features and the state-space model to effectively capture temporal dependencies, the proposed framework models AVs executing collision avoidance strategies while simultaneously prompting HDVs to adapt their behaviors to mitigate potential risks. Results showed that MA-ASS-DDPG demonstrated superior performance in learning maneuvers of both the AV and the FHDV, outperforming counterpart models. Further, the MA-ASS-DDPG was used to reconstruct evasive trajectories of AVs and HDVs in safety-critical scenarios, and the reconstructed data successfully replicated reaction times comparable to real-world observations, further validating the model's effectiveness. Analysis showed that AVs following HDVs reacted 0.3473 s faster than HDV-HDV pairs, while HDVs following AVs reacted 0.2143 s faster, demonstrating more cautious and adaptive driving in response to AV maneuvers. Counterfactual analysis revealed that HDVs following AVs adopt more conservative speeds and larger acceleration variability. In addition, incorporating a safety term into the reward function of the learning framework leads to substantial improvements in safety performance, including reduced conflict occurrences, fewer high-risk deceleration events, and enhanced car-following stability. These outcomes of this study can support safety-aware traffic simulation, scenario-based safety testing, and enhanced AV control strategies in mixed-traffic environments.

Autonomous driving algorithms struggle to achieve sufficient coverage of long-tail scenarios in complex traffic environments, primarily due to the scarcity of high-risk samples in real-world data. Existing scenario generation methods also have limitations, as they mostly rely on trajectory perturbation without realistic perception support. To address this issue, we propose the HSPG (Hazardous Scenario Proactive Generation) framework, a proactive hazardous scenario generation approach based on naturalistic driving data. HSPG systematically amplifies potential risks through structural perturbations of original traffic scenarios. A sliding-window-based risk index is introduced to automatically identify interaction-intensive periods and extract candidate scenarios. A high-risk vehicle detection mechanism then selects critical surrounding vehicles as interaction agents. By integrating a Linear Quadratic Regulator (LQR) with Recurrent Posterior Policy Optimization (RPPO) and adversarial strategies, high-risk trajectories are generated. These trajectories are further transformed into realistic street scenarios via an image synthesis module coupled with real-world map data, forming a comprehensive safety-critical test dataset. Experimental results demonstrate that HSPG effectively identifies latent risks, enhances collision likelihood by at least an order of magnitude under autonomous driving test models, and generalizes across diverse scenarios. A dataset comprising 150 scenarios, 6019 samples, and six multi-perspective camera views has been constructed, providing a valuable benchmark for safety evaluation in autonomous driving. Our dataset can be found at https://huggingface.co/datasets/gitchee/nuScenes-Atk.

Pedestrian automatic emergency braking systems (P-AEB) have recently been introduced in the vehicle fleet to reduce vehicle-to-pedestrian collisions. However, studies on the real-world efficacy of these systems have yielded mixed results. To better understand the factors that influence P-AEB performance, previous simulation and counterfactual studies have evaluated the effects of different P-AEB characteristics on collision avoidance. Previous studies have focused on using a hypothetical P-AEB response model to either estimate the potential benefit of P-AEB or evaluate system configuration performance to optimize P-AEB design. This study aimed to understand the shortcomings of current production P-AEB systems for consumer testing organizations to use for encouraging the continuous improvement of those systems. The present study re-simulated 64 vehicle-to-pedestrian collision cases included in the in-depth Vulnerable Road Users Injury Prevention Alliance database to evaluate the stochastic response of rating-specific P-AEB systems and identify the most challenging pedestrian scenarios and the factors limiting P-AEB performance. Our P-AEB models represented the test responses of systems rated as superior, advanced, or basic by the Insurance Institute for Highway Safety (IIHS). We explored the effects of detection range, detection angle, and the lateral distance threshold for system activation. Results indicated a clear correlation between collision avoidance and the IIHS P-AEB rating. The study also identified three challenging scenarios: (1) highly obstructed cases, (2) high-speed vehicle cases, and (3) cases with high pedestrian crossing speed. None of the explored system designs were able to eliminate collisions in highly obstructed cases due to the late appearance of the pedestrian. In high-speed vehicle cases and in those with high pedestrian crossing speeds, P-AEB performance was limited by the detection range and the lateral distance threshold, respectively. Consumer testing organizations can use these findings to revise existing test programs, improve program relevance for vehicle-to-pedestrian crashes, and incentivize improvements to P-AEB systems.

Child pedestrian casualties in traffic accidents remains high, particularly when they cross the street alone. One contributing factor is their limited ability to identify potential risks. While vehicle motion cues and environmental factors are known to influence hazard perception, a driver's distracted state may also signal risk. However, it remains unclear whether pedestrians, especially children, can assess danger based on a driver's distraction. This study aims to investigate the effects of driver distraction on hazard perception of child (6-10 years old) and adult pedestrians. Participants assessed the safety of crossing at a crosswalk based on videos of approaching vehicles with drivers in various states of distraction (undistracted, texting, chatting, etc.). Results from Experiment 1 show that although both children and adults perceived greater danger when drivers were distracted, children were not as sensitive to different driver states as adults. However, when participants were guided to focus more on driver cues by enlarging driver images (Experiment 2), the influence of driver's distraction on safety assessments increased significantly, particularly for children. This study reveals that even children can perceive potential hazards from driver, which highlights the significant role of driver distraction in pedestrians' safety judgments and provide valuable insights for designing training programs to enhance children's hazard perception skills.

This study proposes a two-step, spatially structured Empirical Bayes (EB) framework for evaluating the safety effectiveness of network-wide countermeasures, leveraging the Network Process Convolution (NPC) model. A central challenge in road safety evaluation is not only estimating treatment effects but also accurately quantifying uncertainty, particularly when interventions generate local and spillover effects. The NPC uses a network-based Gaussian Process with reweighted kernel convolution to capture spatial correlations of collisions along road networks, enabling robust estimation of both site-specific and network-wide effects. The two-step procedure ensures an unbiased prior structure for generating counterfactual outcomes. We conducted a simulation study under varying spatial correlation scenarios and applied the method to the City of Edmonton's Driver Feedback Sign (DFS) program using 10 years of collision data across 1,366 road segments. Performance was benchmarked against the traditional EB Poisson-Gamma (EB-PG) method. Simulations show that while both methods accurately recover counterfactual collisions and reduction ratios, EB-NPC provides more reliable and well-calibrated uncertainty quantification, particularly under moderate to strong spatial correlation. In the Edmonton case study, EB-NPC mostly produced slightly higher estimated reductions and more informative predictive uncertainty, whereas EB-PG remained more robust in areas with weak spatial structure. Beyond numerical estimation, EB-NPC generates continuous spatial risk surfaces, allowing practitioners to visualize network-wide safety patterns and prioritize high-risk segments. Overall, the proposed approach improves recovery of counterfactual outcomes and delivers accurate, interpretable uncertainty characterization, offering a powerful tool for data-driven transportation safety management.

String stability and traffic safety have both received considerable attention in transportation research. However, the analytical relationship between them remains insufficiently explored. This paper addresses this gap by examining four questions: (1) Does string stable traffic imply safe traffic? (2) Does string unstable traffic imply unsafe traffic? (3) Does unsafe traffic imply string unstable traffic? (4) Does safe traffic imply string stable traffic? To investigate these questions, various string stability criteria for both homogeneous traffic and heterogeneous traffic are revisited. Traffic safety is quantified using the Time-To-Collision (TTC) metric, and its connection to string stability is examined through linear stability analysis. The analytical relationships between string stability and safety are derived by incrementally applying conditions from progressing from heterogeneous to homogeneous traffic, providing theoretical answers to the posed questions. The derived relationships are further validated through extensive simulation experiments based on the car-following model calibrated with real-world trajectory data.

Traditional road safety analysis primarily relies on historical crash data, which require long accumulation periods and are constrained by limitations such as insufficient data volume, imprecise location information, and underreporting, potentially leading to biased or delayed assessments of road safety risks. The emergence of connected vehicle (CV) technology provides new opportunities for more timely safety analysis. CVs are equipped with onboard sensors that monitor driving behavior and issue critical warnings, including headway monitoring warnings (HMWs) and forward collision warnings (FCWs). This study aims to proactively identify high-risk expressway segments using CV warning data. Accordingly, an integrated framework is developed, combining spatial hotspot identification and statistical regression modeling. Based on CV data from nine expressways in Shanghai, warning hotspots are identified using Moran's I and Getis-Ord Gi*, indicating locations with spatial clustering of HMWs and FCWs. The relationship between warning frequency and the number of collisions is examined through Poisson and Negative Binomial models estimated with and without incorporating CV warning frequencies as explanatory variables. To address the potential endogeneity between traffic conflicts and collisions, an instrumental variable Poisson model is further employed. The results confirm that HMW and FCW frequencies are positively associated with collisions, and that accounting for endogeneity improves estimation robustness. In addition, hotspot co-occurrence analysis and statistical testing reveal that segments identified exclusively as CV warning hotspots still experience significantly more collisions compared to segments identified as neither warning nor collision hotspots. This suggests that CV warning data can support early detection of emerging safety risks. This study contributes a structured and empirically supported framework that advances the application of connected vehicle data in proactive traffic risk assessment.

Persistent risky behaviours continue to undermine progress reducing traffic fatalities in the United States. Despite the implementation of programmatic interventions combined with increasing awareness, alcohol-impaired driving remains a contributing factor, accounting for nearly 32% of crash deaths in 2022. This study analyzed longitudinal trends in self-reported alcohol-involved driving behaviours, attitudes, and injury outcomes from 2015 to 2024 using Road Safety Monitoring survey data, supported by national fatality statistics from the Fatality Analysis Reporting System. The objective was to identify demographic and behavioural predictors of high-risk outcomes and track patterns over time, especially during periods of disruption. The application of logistic regression models to the survey data showed males were 2.1 times more likely to engage in impaired driving than females and drivers with multiple traffic tickets were more than tenfold likely to do so. The percentage of individuals reporting driving over the legal alcohol limit among those who consumed any level of alcohol in the past 12 months (19,173 out of 26,639 in total) increased from 8.83% in 2015 to 23.99% in 2024. Meanwhile, ride-share use to avoid impaired driving rose from 18.7% in 2016 to 46.4% in 2024. Results indicate a troubling pattern: self-reported alcohol-impaired driving increased significantly between 2015 and 2021 and remained elevated in subsequent years. High levels of public concern about alcohol- and cannabis-impaired driving, along with rising traffic-related injuries, underscore the urgent need for targeted prevention strategies that align with Vision Zero and Safe System goals by addressing the behaviours and groups most at risk.

Time pressure can impair the cognitive functioning of excavator operators, thereby increasing unsafe behaviors and elevating the likelihood of accidents. This study uses a controlled excavator operation task with synchronized behavioral observation and electroencephalography (EEG) recording to examine how escalating time pressure alters operators' cognitive states and safety performance. Results show that as the task deadline approaches, the frequency of unsafe behaviors increases significantly, accompanied by heightened beta-band power and elevated engagement index, reflecting potential cognitive overload under time pressure. To facilitate timely identification of these risk-related neural patterns, we develop RCF-IncepLite, a lightweight EEG-based classification model designed for resource-constrained environments. The model achieves 82.3% accuracy while maintaining minimal computational demands, underscoring its potential for future integration into wearable neuro-sensing systems for early warning of unsafe behaviors. This study provides empirical evidence of the cognitive pathways through which time pressure elevates behavioral risk in construction, and offers a practical methodological foundation for advancing proactive accident prevention in fast-paced construction environments.

Extreme value theory has been receiving much attention of late for proactively estimating crash risk through a two-step procedure that first samples extreme traffic conflicts and then estimates crash risk based on those sampled extremes. Although the existing body of research has encapsulated sampling methods within a predominant conventional technique, there is no universally accepted practice on how to efficiently select threshold values, nor on how to evaluate the sampled extreme conflicts alignment with the conceptual crash severity level framework. This research aims to address these issues by employing machine learning-based sampling methods, which do not require predefined thresholds, and by comparing the sampled extremes with the conceptual severity levels, to assess their alignment. After a review of recent developments in machine learning techniques in transportation and other engineering fields, two promising machine learning sampling models, autoencoder neural network and isolation forest, were investigated using a database of vehicle-to-pedestrian conflicts at urban signalized intersections. Sampled extreme conflicts using the machine learning and conventional sampling techniques-as a baseline -were assessed and compared using two criteria: their visual alignment with the conceptual severity level framework, and their compatibility with the extreme value distribution. The results demonstrate that the extreme conflicts selected based on the machine learning methods better mirror the conceptual severity levels than the conventional sampling technique. Moreover, extremes classified by the isolation forest more closely preserve the characteristics of the empirical tail distributions, demonstrating a better contextual representation for modeling with the extreme value distribution compared to the autoencoder neural network and conventional sampling methods.