Personalized lane departure warning based on non-stationary crossformer and kernel density estimation

The primary causes leading to a high false warning rate in Lane Departure Warning (LDW) systems are the inaccuracy in departure trajectory prediction and inadequate consideration in determining the departure threshold. This study proposes a personalized LDW algorithm based on Non-stationary Crossformer and kernel density estimation. Non-stationary Crossformer is used to predict the future departure trajectory. The model adequately considers the cross-dimension de-pendency, preserving the departure trajectory's non-stationary characteristics while capturing variable information across different time scales. Then the kernel density estimation (KDE) is used to establish a personalized departure threshold for each driver, considering the variance of risk tolerance for departure among drivers. Leveraged by the impact of each driver's historical departure area, the KDE divides the historical area into two regions and determines the threshold. Vali-dation is carried out using Shanghai natural driving data collected from 10 drivers. The results show that, compared with baseline modes, Non-stationary Crossformer can accurately predict departure trajectory when there are significant trajectory changes, and the KDE can better deter-mine the personalized departure threshold. The proposed LDW algorithm reduces the false warning rate to 2.4%, which is 2.4 %-24 % lower than Baseline models. The algorithm would con-tribute to improving the LDW in the future.