An Analysis and Identification Methods of Dangerous Driving Behavior Characteristics at Highway Tunnel Entrances and Exits

Dangerous driving behaviors frequently occur at highway tunnel entrances and exits, posing a high risk of traffic accidents. To address the challenge of ineffective driving risk assessment caused by the inability to continuously monitor trajectory data at tunnel transition zones, this study designs a radar-video fusion trajectory sampling system with a monitoring range covering 250 meters inside and outside the tunnel portal. A dangerous driving behavior identification method based on feature parameter optimization is proposed. Based on trajectory data at tunnel entrances and exits, the characteristics of driving behavior in these zones are analyzed, and four types of dangerous driving behaviors including sudden acceleration or deceleration, serpentine driving, high-risk car-following, and aggressive lane-changing, are selected to construct a dangerous driving behavior spectrum. A risk quantification method is used to measure indicators of the four dangerous driving behaviors, and the interquartile range (IQR) method is applied to set threshold boundaries for the feature parameters. Based on these thresholds, driving risk points exceeding the boundary values are identified and visualized, and the spatial distribution characteristics of the four types of dangerous driving behaviors are preliminarily obtained. To balance the dataset, random oversampling (ROS), synthetic minority oversampling technique (SMOTE), and adaptive synthetic sampling (ADASYN) are used for sample preprocessing. Three ensemble learning methods: eXtreme gradient boosting (XGBoost), light gradient boosting machine (LGBM), and adaptive boosting (AdaBoost), are orthogonally combined with the above sampling methods to construct balanced-ensemble coupled algorithms. A total of 12 dangerous driving behavior recognition models are established, including those based on single ensemble learning algorithms and orthogonally combined balanced-ensemble algorithms. The performance differences among various models are validated through model testing to determine the optimal recognition model. Spearman correlation analysis is employed to identify key parameters and enhance model recognition performance. The research results indicate that due to the complex traffic environment and fluctuating driver behaviors, highway tunnel entrances and exits are high-risk zones for traffic accidents. Among the three single-modality ensemble models and nine balanced-ensemble coupled models evaluated, the SMOTE-LGBM coupled model based on sample optimization demonstrates superior recognition performance for dangerous driving behaviors in tunnel transition zones. Its precision, F-score, and AUC values range from 91.2% to 91.4%, 0.913 to 0.918, and 0.907 to 0.912, respectively, outperforming other algorithms and maintaining consistently high levels.