Utilizing YOLOv8 for enhanced traffic monitoring in intelligent transportation systems (ITS) applications

Abstract

The increasing demand for artificial intelligence-based motor vehicle detection in Intelligent Transportation Systems (ITS) applications highlights the significance of advancements in this field. The introduction of YOLOv8, the latest iteration in the YOLO algorithm series, presents a new avenue for exploring the potential of this detection algorithm within the ITS domain. The algorithm has not been previously tested in applications such as vehicle detection, which highlights a gap in the existing literature. This presents an opportunity to explore its capabilities and contributions in traffic monitoring and vehicle detection. This study aims to address this gap by employing YOLOv8 for vehicle detection within the broader context of ITS applications. Distinguishing itself from its predecessors, YOLOv8 features a decoupled head structure and employs a C2f module instead of C3. Extensive testing was performed using datasets acquired through aerial monitoring with a drone. Special emphasis was placed on ensuring a diverse array of target objects during dataset creation, a detail frequently neglected in comparable studies. The algorithm's training not only facilitated an evaluation of its ability to generalize and process data proficiently but also provided initial insights into its potential for real-time applications. The model underwent a comprehensive series of performance tests, revealing both strengths and weaknesses and outlining its capabilities and limitations. In a comparative analysis, the study systematically compared the performance metrics of YOLOv8 with those of YOLOv5, a widely adopted model in ITS research. Precision assessments unveiled a significant disparity, with YOLOv8 exhibiting an 18% increase in precision compared to YOLOv5. Further investigation into the inference times of both algorithms highlighted the superior processing speed performance of YOLOv8. The study's findings shed light on the limitations of the detection process, attributing misclassifications to factors such as variations in vehicle shapes, lighting conditions, and relative sizes.