Risky Driving Behavior Recognition Based on Vehicle Trajectory.

Area of Science:

• Transportation Safety
• Computer Vision
• Artificial Intelligence

Background:

• Risky driving behaviors pose significant safety challenges on roadways.
• Existing methods for detecting dangerous driving often lack robustness or require specific sensor data.
• Automated recognition of driving behavior from surveillance video is crucial for intelligent transportation systems.

Purpose of the Study:

• To propose a novel Measurement of Risk (MOR) method for recognizing specific risky driving behaviors.
• To develop and validate an MOR-based model for quantifying collision risk.
• To establish effective methods for selecting MOR thresholds to accurately classify risky driving.

Main Methods:

• Trajectory data extraction from surveillance videos (UAV footage used).
• Development of an MOR-based risk evaluation model for speed-unstable, serpentine, and car-following driving.
• Implementation of distribution-based and boxplot-based methods for MOR threshold selection.
• Consideration of vehicle types (cars vs. heavy trucks) in the analysis.

Main Results:

• The MOR method successfully quantifies collision risk for identified risky driving behaviors.
• Significant differences in MOR thresholds were observed between cars and heavy trucks for certain behaviors.
• The proposed methods achieved high recognition accuracies: 91% (boxplot) and 86% (distribution).
• Minimal difference (<3.5%) in recognized risky behavior proportion between training and testing datasets.

Conclusions:

• The proposed MOR method provides a validated approach for recognizing risky driving behavior from video surveillance.
• The model demonstrates effectiveness across different vehicle types and driving scenarios.
• This approach has broad applicability in enhancing road safety through intelligent video analysis.