AdaBoost-Based Identification Model of Risky Lane-Changing Behavior

Unsafe lane-changing behaviors can lead to serious traffic accidents in weaving areas. This research proposes an end-to-end model for the identification of risky lane-changing behaviors. A customized Darknet architecture is utilized to obtain vehicle trajectories. Following this, a novel lane-change data extraction framework is instituted, considering lateral deviations of vehicles. Subsequently, two traffic conflict indices are integrated to evaluate the safety implications. Finally, AdaBoost is adopted to realize the identification of risky lane-changing behaviors. For validation, the model is evaluated using Chinese drone data, NGSIM, and HighD. A feature selection methodology is employed to identify optimal discriminative criteria, while comparisons are undertaken with Random Forests and Support Vector Machines. The AdaBoost exhibits superior performance, achieving a maximal precision and recall rate of 87.04% and 87.50%. This result is crucial for designing driver assistance systems, managing accident hotspots, and responding quickly to major lane-changing incidents.