Interaction patterns exploration and risk assessment within lane changing in terms of initial scenarios and evolution...

The lane-changing (LC) process is a sequential interaction process influenced by multiple factors, including the driver's decision-making, speed, traffic density, road conditions, and the behavior of surrounding vehicles. However, it remains unclear in which scenarios drivers tend to execute lane changes and how the lane changing vehicle and the surrounding vehicles jointly complete the entire lane-changing process in these scenarios. This research establishes a two-level integrated clustering framework tailored to the specifics of LC interactions. At the upper level, the normalized k-means algorithm and pairwise aggregation SSE (sum of the squared errors) are jointly used in this research to handle high-dimensional data and analyze the scenarios in which drivers tend to execute lane changes (i.e., LC initial-scenario). At the lower level, the interaction feature extraction method is utilized to explore the LC interaction patterns corresponding to different LC initial-scenarios. Based on the highD trajectory data and the aforementioned two-level integrated clustering framework, this research identifies 9 major LC initial scenarios and 20 major LC interaction patterns of human drivers, and establishes the relationships between different LC initial scenarios and LC interaction patterns. Particularly, we discover that the interaction patterns could be further categorized into global general pattern, local general pattern, and special pattern. Additionally, by introducing the driving risk quantification indicators, this research identifies risky and safe LC initial scenarios and interaction patterns. The findings offer an in-depth understanding of the entire lane-changing process, providing a theoretical foundation for autonomous driving tests, lane-changing collaborative design, and traffic warning systems.