Enhanced Vehicle Classification via Axle-Based Deep Learning Models

In an increasingly automated world, the automation of toll gates has not eliminated the challenges associated with vehicle classification, primarily due to vulnerabilities in sensor technology. These challenges result in significant inaccuracies in toll rate determination. This research addresses this issue by proposing the implementation of computer vision techniques to classify vehicles according to their axle count, a method that remains underexplored within the domain of deep learning. The study focuses on the YOLO (You Only Look Once) algorithm due to its advantageous speed, accuracy, and user-friendly application within Convolutional Neural Networks (CNNs). The primary objective of this paper is to evaluate and compare the performance of four YOLO models—YOLOv5, YOLOv6, YOLOv7, and YOLOv8—to identify the most effective model for vehicle classification. Experimental results indicate that the YOLOv8 model achieves the highest performance metrics, with an accuracy of 0.954, a recall score of 0.972, a mean Average Precision (mAP) at 50 of 0.987, and an mAP across 50-90 of 0.891. These findings demonstrate that the YOLOv8 model significantly outperforms its counterparts in vehicle detection and axle counting tasks. It is our hope that this research will contribute to future advancements in the automatic classification of various vehicle types.