Hybrid feature selection for real-time road surface classification on low-end hardware: A machine learning approach

Accurately classifying road surface conditions is crucial for maintaining road safety and facilitating effective maintenance. Vibration-based methods have proven effective in this field, utilizing vehicle vibration patterns to determine road surface conditions. One of the challenges in this field is using optimal datasets and classification models that meet real-time applications on low-end hardware devices. This paper proposed a hybrid filter-wrapper algorithm for feature selection based on an index of important features used in the pavement classification task; it takes advantage of two pure methods: the fast speed of the filter and the efficiency of the wrapper method. The influence of velocity on the collected data of inertial sensors is also evaluated, and the number of sensors in the pavement classification task is also considered. We evaluate the proposed method in three machine learning models, Random Forest, Gradient Boosting, and XGBoost, to classify road surface type into three classes: asphalt, dirt, and cobblestone. The proposed feature selection approach significantly improves model efficiency and execution speed while maintaining high accuracy. This study highlights the important role of inertial sensor data collection, pre-processing, and feature selection in enhancing model performance for pavement classification, demonstrating its value in both computational efficiency and accuracy. • A hybrid feature selection method is proposed for road surface classification. • The method uses inertial sensor data and tree-based machine learning models. • Classification accuracy remains high even with reduced feature sets. • Achieves 94% accuracy on XGBoost with only 30 out of 126 features. • Suitable for real-time use on low-end hardware.