Analysis and improve of the YOLOv8 object detection model performance using a custom BDD100K dataset

Abstract

This study aimed to evaluate the performance of the YOLOv8 model in detecting vehicles and traffic elements within complex urban environments, using a customized version of the BDD100K database containing nine major traffic categories. Both YOLOv8n and YOLOv8m were initially trained using standard settings. We then proposed a methodology to improve the YOLOv8m model by incorporating advanced data augmentation techniques, increasing image resolution, and changing the optimization algorithm to AdamW.The results showed that the YOLOv8m model outperformed YOLOv8n in all evaluation indicators, achieving mAP50 = 0.5905 versus 0.4219, and Precision = 0.6801 versus 0.6288. After implementing the proposed methodology, YOLOv8m achieved further performance improvements, reaching mAP50 = 0.6116 and Precision = 0.7026, along with a decrease in loss values ​​on the validation data. This indicates increased model stability and improved generalizability. These results reflect the importance of data allocation and improving training strategies to increase the accuracy of detection models in intelligent transportation systems.