Real-time safety analysis is the premise and foundation of Active Traffic Management (ATM). Traffic safety level is evaluated in a real-time manner. Once crash-prone conditions are identified, proactive traffic management can be triggered to reduce crash risk and prevent traffic accidents. As the lifeline of road transportation, expressways have high operating speeds, large traffic flow, and the high-level mixing of large vehicles, which pose significant challenges to driving safety. Improving expressway safety is an urgent need. In recent years, the development of smart expressways and intelligent connected vehicles has put forward new requirements for traffic safety. The concept of safety management has shifted from reactive improvement to proactive prevention, and real-time safety analysis has become a focus and research frontier in the field of road safety. Active safety management of traffic flow requires a more extensive and precise real-time safety evaluation method that covers a wider range of risk scenarios.

However, there are many challenges in developing real-time safety models for expressways. This study summarizes three key technical issues based on previous research: (1) How can a real-time safety model be developed without historical accident data? (2) How to identify high-risk scenarios covering the entire evolution process of crash risk? (3) How to understand fine-grained crash risk mechanisms and improve real-time safety assessment from minute-level to second-level? To address these research challenges, this study proposes a real-time expressways safety analysis method based on traffic conflicts theory. The research content can be summarized into four aspects: analysis of high-frequency traffic conflict conditions for practical applications; identifying conflict precursors and investigating the heterogeneous effects of real-time traffic flow features on conflict occurrence; real-time safety assessment and conflict mechanism analysis of continuous traffic flow towards a connected environment; and real-time conflict prediction considering vehicle platoon features under limited data supply. The above four studies have gradually improved real-time safety analysis from the "minute level" to the "second level" in the temporal dimension, from "road sections" to "vehicle platoon" in the spatial dimension and have gradually transitioned the application environment from "traditional expressway" to "cooperative vehicle infrastructure system". The specific framework of the research is summarized as follows:

Firstly, A novel real-time safety analysis framework driven by high-resolution trajectory data is developed, which integrates the concept of aggregate safety analysis and traditional real-time safety analysis extending the risk measurement from the probability of crash occurrence to the frequency of conflicts. To ensure practicality, a virtual detector algorithm is designed to extract traffic flow characteristics of expressway cross-sections, such as upstream traffic flow volume, upstream speed variation coefficient, and the proportion of large vehicles in upstream traffic, from trajectory data sets. A random effect negative binomial model is developed to identify contributory factors that affect conflict frequency.

Secondly, potential precursors before conflict occurrence are identified based on the significant correlation between traffic conflict and dynamic traffic flow. To address the problem of coarse time granularity in real-time safety research, fine-grained temporal units for conflict occurrence are applied. Conflict analysis is refined from the minute level to the second level. A research workflow is developed from knowledge mining to predictive application by a two-stage analysis framework composed of statistical models and ensemble learning models. A random parameter model with heterogeneity in means is used to investigate the heterogeneity effects of fine-grained traffic flow on conflict occurrence and identify the real-time precursors that affect the probability of conflict. The results help give insights into the conflict mechanism from the perspective of mesoscopic traffic flow

Then, a collaborative work mode for real-time expressway safety assessment is developed towards a connected environment with the integrated use of multiple sources of data from roadside detectors and connected vehicles. Ensemble learning models are employed to assess real-time safety under collaborative work mode. Explainable machine learning indicates that factors such as speed fluctuation and the proportion of large vehicles are positively correlated with conflict occurrence in continuous traffic flow. Furthermore, experiments on the market penetration rates of connected vehicles (CV) are conducted to explore the model performance in different levels of connected environments. Experimental results show that as the market penetration rate (MPR) of CV increased, the model performance is gradually improved. Even in the early stages of the connected environment, namely, the low penetration rate situation, the collaborative mode could significantly improve the performance of real-time safety model. A moderate MPR could ensure satisfactory performance for real-time safety models. To the best of our knowledge, this is the first time that investigates the collaborative work mode for real-time safety models with the consideration of CV MPR. Based on the in-depth exploration of the model performance gains and change patterns of models in a connected environment, this study can provide a solid reference for the development of real-time safety models in future mixed traffic environments with both traditional and connected vehicles.

Finally, the study focuses on the risk averaging problem and sight distance limitation of current safety research. A "beyond-line-of-sight" risk perception approach is proposed to capture vehicle platoon risks beyond the workable range of onboard equipment. The deep generative model is applied to obtain reliable conflict data augmentation with limited conflict data input. The results demonstrate that the Wasserstein generative adversarial network with gradient penalty can capture the distribution pattern of conflicts and derive high-risk crash scenarios. Synthetic conflict data has similar data distribution to real conflict data, which can help overcome the inevitable problem of imbalanced data in the development of real-time safety models. A vehicle-level conflict prediction model is then established using traffic characteristics of vehicle platoons. The dual-stream neural network model trained with the rebalanced dataset outperforms other models in all evaluative metrics. The proposed method can enhance the risk perception for autonomous vehicles beyond “sight range”. The data augmentation technique with limited conflict data has great potential to enrich the test scenarios for autonomous driving.

Traffic conflicts can effectively address the limitations of new expressways or segments with no historical accidents, avoiding inherent data quality issues of crash data. Additionally, risk measurement based on crash probability cannot cover all high-risk scenarios. The results of this study can further help understand the risk mechanism and enrich the risk measurement. The analytical method proposed in this study can be flexibly deployed, and the real-time safety evaluation of the "expressway segment-vehicle platoon" will provide a foundation for active safety management and safety warning via a vehicle-infrastructure cooperative system. It can also provide driving style references for the safe motion planning of autonomous vehicles, further enhancing expressway safety in future intelligent connected environments.