Performance Analysis of Yolo Versions on Road Environment

Objectives: This study evaluates the performance of YOLO object detection models, from YOLOv5 to YOLOv10, in detecting vehicles, pedestrians, and traffic signs in road environments, with a focus on their applicability to real-time autonomous driving. Theoretical Framework: The research builds on the evolution of YOLO as a leading single-stage object detection framework, which is recognized for balancing accuracy and speed compared to traditional two-stage methods. Method: A custom dataset of 5,950 annotated images collected under diverse weather and illumination conditions was used to train and evaluate the models. Performance was assessed using precision, recall, F1-score, mean Average Precision (mAP), and inference time, with preprocessing carried out in Roboflow and experiments executed on Google Colab with GPU support. Results and Discussion: The results show that YOLOv5x achieved the highest accuracy, with a mAP of 76.2% and an F1-score of 77.34%, while YOLOv8n demonstrated the fastest inference time at 2.2 milliseconds. Persistent challenges were identified in detecting small objects and addressing dataset class imbalance, which significantly affected performance. Research Implications: The findings highlight the need for dataset balancing and careful model selection to improve detection in autonomous driving applications. Originality/Value: This study offers a comprehensive comparison of YOLO versions in urban road environments, providing valuable insights for researchers and practitioners in intelligent transportation systems.