Refined conflict prediction: improving traffic safety and efficiency simultaneously by fully considering the benefits of connected vehicles

An innovative approach has been investigated by researchers in recent years which utilizes real vehicle trajectory data to simultaneously obtain traffic state variables and corresponding conflict data for conflict prediction. Furthermore, considering the future popularity of connected vehicles (CVs), promising more detailed traffic information, related studies have also been explored. However, existing relevant studies mainly focused on predicting lane-level conflicts across the roadway, which may not fully utilize the available roadway space, thus potentially compromising traffic efficiency. Therefore, there is a need for a more refined prediction of conflict that can be achieved when the advantages of CVs are fully considered. In this study, data from HighD were applied, and lanes (i.e. inner, middle and outer lanes) were further segmented into smaller sub-segments for refined analysis. Subsequently, conflict prediction was performed for the overall roadway as well as for these individual sub-segments based on machine learning and deep learning models. Additionally, this study explored conflict prediction when considering CVs with varying market penetration rates (MPR) and distribution patterns (i.e. uniform distribution (UD) and random distribution (RD)), as well as their combined influence. Results indicate that predictions for the sub-segments show better performance compared to those for the overall roadway, highlighting the effectiveness of the proposed approach for refined prediction. Besides, the Bidirectional Long Short-Term Memory (Bi-LSTM) consistently exhibits superior performance in predicting conflicts among the selected models. Moreover, the prediction performance becomes better when CVs are considered and an improvement can be observed as the MPR of CVs increases, particularly when CVs are under the uniform distribution. To verify the reliability of these results, sensitivity analysis was further carried out and the results were consistent with the previous findings. Findings of this study contribute to the refined prediction of real-time conflicts and development of proactive safety management in the future.