Boosting Vehicle Classification with Augmentation Techniques across Multiple YOLO Versions

Abstract

In recent years, computer vision has experienced a surge in applications across various domains, including product and quality inspection, automatic surveillance, and robotics. This study proposes techniques to enhance vehicle object detection and classification using augmentation methods based on the YOLO (You Only Look Once) network. The primary objective of the trained model is to generate a local vehicle detection system for Malaysia which have the capacity to detect vehicles manufactured in Malaysia, adapt to the specific environmental factors in Malaysia, and accommodate varying lighting conditions prevalent in Malaysia. The dataset used for this paper to develop and evaluate the proposed system was provided by a highway company, which captured a comprehensive top-down view of the highway using a surveillance camera. Rigorous manual annotation was employed to ensure accurate annotations within the dataset. Various image augmentation techniques were also applied to enhance the dataset's diversity and improve the system's robustness. Experiments were conducted using different versions of the YOLO network, such as YOLOv5, YOLOv6, YOLOv7, and YOLOv8, each with varying hyperparameter settings. These experiments aimed to identify the optimal configuration for the given dataset. The experimental results demonstrated the superiority of YOLOv8 over other YOLO versions, achieving an impressive mean average precision of 97.9% for vehicle detection. Moreover, data augmentation effectively solves the issues of overfitting and data imbalance while providing diverse perspectives in the dataset. Future research can focus on optimizing computational efficiency for real-time applications and large-scale deployments.