Transformer-Based Classification of Road Conditions Using Vehicular Sensor Data
Ibrahim Aslam, Sazia Mahfuz · Procedia Computer Science · 2025
Precise classification of road surface types is a crucial aspect of current advanced transport systems due to its importance in safety, routing, and prognostic maintenance. This work proposes a transformer model to predict the road surface types (paved or unpaved) based on vehicular sensor data derived from several onboard sensors such as accelerometers, gyroscopes, magnetometers, and temperature sensors. The methodology involves a feature selection step based on the Random Forest classifier ranking without negatively affecting the predictive accuracy. An initial transformer design was furthermore developed with multi-head attention, which was trained and validated on time-series data with regularization. The proposed model achieved a weighted average F1-score of 0.97, which supports the use of transformers in analyzing vehicular sensors and their application in the field of smart transportation.