Despite the technological advances in automated driving systems, traffic accidents involving automated vehicles (AVs) continue to occur, raising concerns over the safety and reliability of automated driving. For the smooth commercialization of AVs, it is necessary to systematically assess the driving safety of AVs under the various situations that they potentially face. In this context, these various situations are mostly implemented by using systematically developed scenarios. In accordance with this need, a scenario generation framework for the assessment of the driving safety of AVs is proposed by this study. The proposed framework provides a unified form of assessment with key components for each scenario stage to facilitate systematization and objectivity. The performance of the driving safety assessment scenarios generated within the proposed framework was verified. Traffic accident report data were used for verification, and the usefulness of the proposed framework was confirmed by generating a set of scenarios, ranging from functional scenarios to test cases. The scenario generation framework proposed in this study can be used to provide sustainable scenarios. In addition, from this, it is possible to create assessment scenarios for all road types and various assessment spaces, such as simulations, proving grounds, and real roads.

For consumers to have confidence in the safety of automated vehicles (AVs), AVs must be assessed using systematically developed scenarios to verify driving safety and reliability. In particular, verification using scenarios has been widely performed for the assessment and certification of AVs. This study aims to develop test cases based on driving trajectories to assess the driving safety of AVs. To achieve this, concrete scenarios were systematically developed from functional and logical scenarios. Drone video data analysis was conducted to extract representative lane-change trajectories for AVs on expressway ramp sections. Subsequently, the test cases were selected from concrete scenarios through simulations using time-to-steer (TTS). Finally, the effectiveness of utilizing trajectories for scenario-based driving safety assessments was verified. Furthermore, it is expected that this approach can be applied to other driving patterns by providing a detailed procedure for the test case developed in this study.

Despite advancements in automated driving technology, traffic accidents remain a significant safety concern in urban areas because of complex traffic environments and current technical limitations of automated vehicles (AV). Structured, quantitative scenario-based evaluations are essential for assessing AV safety. This study aims to generate functional scenarios by accurately inferring vehicle and pedestrian behavior information from unstructured traffic accident descriptions. The dataset comprises 137,432 traffic accident cases collected by the Korean National Police Agency. A total of 1,774 training samples were extracted using term frequency–inverse document frequency keyword analysis. We then evaluated the prediction performance of three large language models (LLM): Claude 3, GPT-4o, and GPT-4o with fine-tuning, using 500 randomly selected traffic accident records that lacked keyword-extractable behavior information. GPT-4o with fine-tuning achieved the highest accuracy of 81.1%. Using this model, 26 functional scenarios were generated. The risk levels for each scenario were evaluated by integrating the equivalent property damage only index with risk predictions from logistic regression and XGBoost models. A weighted average of the normalized results was used to rank the scenarios. Scenarios such as “turning left / crossing,” “straight / crossing,” and “turning right / crossing” were consistently identified as high-risk. This study demonstrates the effectiveness of LLM-based contextual inference in converting real-world traffic accident data into functional scenarios, and contributes to the advancement of AV safety evaluation frameworks.

This study aims to identify the factors that influence the occurrence of traffic accidents to improve motorway traffic safety. Various data, including the frequency of traffic accidents, traffic volume, geometric structure, and congestion level, were collected from individual sections of motorways in South Korea. Using the collected data, a traffic accident frequency prediction model was developed by applying an explainable artificial intelligence (AI)-based approach. The developed deep neural network model was combined with Shapley Additive Explanations to identify the variables that significantly affect the frequency of traffic accidents. The analysis identified five significant factors: segment length, total traffic volume, the proportion of truck traffic, the number of dangerous driving behaviors, and the duration of congestion. The results demonstrate the potential of using explainable AI in predicting traffic accident frequency. By identifying the factors that influence traffic accidents using this model, we can pinpoint areas for improvement, which may ultimately help reduce highway traffic accidents.

Traffic information has traditionally been estimated using fixed point-based detection systems for road operation and management. However, this approach faces limitations in spatial and temporal coverage, as well as high maintenance costs. With the expected growth of automated vehicles (AVs), they can be utilized as probe vehicles, providing new opportunities for real-time traffic monitoring. Unlike conventional point-based detection systems, mobile automated probe vehicles (APVs) can collect traffic data even in spatially shadowed areas with a relatively small number of vehicles. This study proposes a methodology for estimating traffic density using individual vehicle information obtained from APVs. In addition, it derives optimized APV operation strategies that account for the APV penetration rate (APP) in mixed traffic, traffic volume, the maximum number of vehicles detectable per APV (MVD), and data collection interval. Based on three key parameters—APP, traffic volume, and MVD—, 312 scenarios were generated using microscopic traffic simulation. For each scenario, traffic density estimation errors were compared with ground-truth values using MAE, MAPE, and RMSE. The results indicate that an APP of 50% and MVD of 2 vehicles provide a balance between estimation accuracy and operational efficiency. In addition, the appropriate data collection interval ranges from 1-minute to 15-minutes. This study proposes an optimized APV operation strategy for traffic density estimation, highlighting its potential to enhance the accuracy, efficiency, and practical applicability of future traffic monitoring systems.