Related Experiment Video
Updated: Aug 5, 2026
07:15
Tactile Vibrating Toolkit and Driving Simulation Platform for Driving-Related Research
Published on: December 18, 2020
ELM-AdaBoost-Based Recognition of Risky Driving Behavior
Dudu Guo1,2, Entong Liu3, Guoliang Chen2,4
1Xinjiang Key Laboratory of Green Construction and Smart Traffic Control of Transportation Infrastructure, Xinjiang University, Urumqi 830017, China.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a fleet-specific method using 90th-percentile (P90) statistics to identify risky driving behaviors in commercial fleets. This adaptive approach enhances road safety management by creating a more accurate dataset for driver behavior analysis.
Area of Science:
- Road safety
- Transportation engineering
- Data science
Background:
- Traditional traffic safety management relies on fixed thresholds, which are inadequate for specialized fleets like hazardous chemical transporters.
- Operational characteristics of specific vehicle fleets, especially under speed supervision, necessitate adaptive safety metrics.
- Existing methods struggle to accurately classify risky driving behaviors in diverse fleet operations.
Purpose of the Study:
- To develop a fleet-specific relative threshold labeling framework for identifying risky driving behaviors.
- To create an adaptive system that overcomes the limitations of fixed empirical thresholds in road safety management.
- To construct a specialized dataset for training and validating machine learning models for risky driving behavior classification.
Main Methods:
- Proposed a labeling framework based on 90th-percentile (P90) statistics to adaptively identify five risky driving behaviors (speeding, rapid acceleration/deceleration, sharp turning, sharp lane changing).
- Extracted risky driving events from 481 GPS trajectory trip segments to construct a dedicated dataset.
- Employed an Extreme Learning Machine combined with Adaptive Boosting (ELM-AdaBoost) as the recognition model for performance validation.
Main Results:
- The P90-labeled dataset enabled an average recognition accuracy of 95.8% for the five identified risky driving behaviors.
- The ELM-AdaBoost model trained on the P90 dataset significantly outperformed other classification algorithms, including Random Forest, BP neural network, and standalone ELM.
- The proposed labeling framework demonstrated its ability to produce a well-separated and learnable dataset for improved driver behavior analysis.
Conclusions:
- The fleet-specific P90 labeling framework effectively addresses the limitations of fixed thresholds in classifying risky driving behaviors.
- The adaptive approach enhances the accuracy and reliability of road traffic safety management systems, particularly for specialized vehicle fleets.
- The developed dataset and recognition model provide a robust foundation for advancing research in intelligent transportation systems and driver safety.