<![CDATA[This study assesses the detection and classification performance of YOLOv5, YOLOv7, YOLOv8, and YOLOv9 for vehicle license plate colors in Indonesia, supporting the electronic ticketing system (e-tilang). The dataset consisted of 1,214 images from video footage captured in Bandar Lampung, comprising five color categories: black, white, yellow, red, and non-plate. The models were trained using transfer learning with COCO pre-trained weights, evaluated using precision, recall, F1-score, mAP50, and mAP50-95, and tested under real-world moderate and crowded traffic conditions. The results show that YOLOv9 consistently outperformed all other models, achieving the highest precision (97.20%), recall (96.50%), F1-score (96.85%), mAP50 (98.10%), and mAP50-95 (80.50%), with the fastest inference time of 6.8 ms per image (approximately 147 FPS). YOLOv8 ranked second, followed by YOLOv7 and YOLOv5. Across all models, the non-plate category remained the most challenging, while white and yellow plates were occasionally misclassified under low-light conditions. In conclusion, YOLOv9 is recommended for deployment in Indonesia's e-tilang system due to its best balance of accuracy and speed. Future work should expand the dataset to more diverse geographical locations, evaluate model performance under extreme weather conditions, and deploy the model on edge devices to validate real-world performance.]]>
Copyrights © 2026
