Using Support Vector Machines to Classify Road Surface Conditions to Promote Safe Driving.

Area of Science:

• Computer Science
• Engineering
• Environmental Science

Background:

• Accurate detection of road surface conditions during winter is critical for traffic safety and efficient road management.
• Existing methods may lack accuracy or generalizability in adverse weather conditions.

Purpose of the Study:

• To develop and evaluate a data-driven learning model for accurate and generalizable estimation of road surface conditions.
• To assess the performance of Support Vector Machine (SVM) models with different kernel functions and classification strategies.

Main Methods:

• Utilized a Support Vector Machine (SVM) machine learning model with linear, Gaussian, and second-order polynomial kernel functions.
• Employed soft margin classification and two learner designs (one-vs-one, one-vs-all) for multi-class classification.
• Calculated posterior probabilities using the sigmoid function to analyze classification confidence.

Main Results:

• Classification errors for most SVM classifiers were below 3%, demonstrating high accuracy in identifying road surface conditions.
• One-vs-one learners exhibited generalization performance within a 4% error rate.
• Posterior probabilities effectively identified atmospheric and road surface conditions associated with hazardous situations.

Conclusions:

• Data-driven learning models, particularly SVM, show significant potential for accurately classifying road surface conditions in adverse winter weather.
• The developed model contributes to improved traffic safety and proactive road management strategies.
• Posterior probability analysis enhances the interpretability and predictive power for hazardous conditions.