Current status of lane change intention recognition of autonomous vehicles

In recent years, significant progress has been made in the research of data-driven automatic driving vehicle lane-changing intention recognition, and researchers have released a large number of research results. In response to some common technical challenges faced in this field, such as identifying lane-changing processes, missing lane-changing labels, and imbalanced data categories, different data-driven methods were used for classification, mainly including traditional machine learning-based, deep learning-based, and ensemble learning-based lane-changing intention recognition methods. The research achievements of these methods in recent years were reviewed and summarized. There are two mainstream options for determining lane-changing behavior, one based on whether to cross the lane line and the other based on other characteristics during the lane-changing process. For vehicles that have not crossed the lane line, early recognition methods based on the driver's intention to change lanes are mainly used, and when the vehicle crosses the lane line, this behavior is usually used to identify the complete lane-changing process. In the study of lane-changing intention annotation, researchers conducted in-depth discussions on the impact of fixed time windows and heading angle thresholds on annotation accuracy. To find the optimal parameters, such as the optimal fixed time window and heading angle threshold, researchers used a grid search for optimization. Although this method performs well in fixed driving scenarios, achieving adaptive parameter adjustment in different driving scenarios remains a challenge. To address the issue of imbalanced data categories in lane changing, researchers adopt two strategies: one is to adjust the data sampling method and use under sampling and oversampling techniques to balance the number of samples in each category; The second is to use classification models with strong adaptability to imbalanced data, such as ensemble learning algorithms or cost-sensitive learning, to maintain good classification performance.