ELM-AdaBoost-Based Recognition of Risky Driving Behavior.

The classification of risky driving behaviors is a key aspect of improving road traffic safety management. A major challenge in this field is that fixed empirical thresholds cannot adequately reflect the operational characteristics of specific vehicle fleets, particularly hazardous chemical transportation vehicles operating under mandatory speed supervision. To address this gap, this study proposes a fleet-specific relative threshold labeling framework based on 90th-percentile (P90) statistics, which adaptively identifies five types of risky driving behaviors-speeding, rapid acceleration, rapid deceleration, sharp turning, and sharp lane changing-from GPS trajectory data. Using this framework, risky driving events were extracted from 481 trip segments and a dedicated risky driving behavior dataset was constructed. To validate the effectiveness of the proposed labeling strategy, an Extreme Learning Machine combined with Adaptive Boosting (ELM-AdaBoost) was employed as the recognition model and compared with commonly used classification algorithms. The results show that the model trained on the P90-labeled dataset achieves an average recognition accuracy of 95.8% for the five behavior categories, outperforming Random Forest, BP neural network, and standalone ELM, thereby confirming that the proposed labeling framework produces a well-separated and learnable dataset.