<![CDATA[Road safety is an important aspect in reducing accident risks, especially for motorcycle riders. To improve compliance with helmet use, this study compares the performance of two deep learning–based object detection models, namely YOLOv8 and RT-DETR, using a Roboflow dataset consisting of 3,735 images with two classes: with helmet and without helmet. The research process includes data acquisition, preprocessing (512×512 pixels), model training conducted in Visual Studio Code using an Nvidia GTX 1070 Ti GPU with the Ultralytics framework (100 epochs, AdamW optimizer, 0.0005 learning rate, 25 patience), testing on images, videos, and real-time inputs using last.pt, as well as evaluation through precision, recall, mAP, and confusion matrix, followed by implementation of the best algorithm in a local Streamlit web application.The results show that RT-DETR achieved slightly better training performance in terms of mAP50–95, while YOLOv8 performed better during real-world testing with more stable accuracy, particularly for the with helmet class. YOLOv8 reached up to 100% accuracy in video and real-time testing, whereas RT-DETR performed better in the without helmet class, achieving 95% accuracy on image data and up to 100% in video testing. Overall, YOLOv8 was selected as the best model for implementation in the Streamlit-based helmet detection application because it is faster, more stable, and more accurate. This system has the potential to support intelligent ETLE enforcement to enhance traffic safety in Indonesia.]]>
Copyrights © 2025
