Development and Performance Evaluation of a Helmet Detection Prototype System Using CIRA CORE

Abstract

This research aimed to design and develop a helmet detection prototype system and evaluate its detection performance within Uttaradit Rajabhat University. The system was developed using CIRA CORE and trained with image datasets consisting of 200, 300, 400, and 500 images. The experimental process included image preparation, helmet annotation, and model training using a deep learning training tool. The detection performance of the developed models was subsequently compared with that of the YOLOv3 model. The experimental results revealed that the models trained with 200 and 300 images were unable to detect helmets. The model trained with 400 images achieved an average accuracy of 13.5%, whereas the model trained with 500 images achieved an average accuracy of 57.9%. In comparison, the YOLOv3 model achieved an average accuracy of 48.3%. Furthermore, camera-angle evaluation showed that at the 45-degree angle, the system successfully detected helmets 9 times, achieving an average accuracy of 57.2%, whereas at the 90-degree angle, the system detected helmets 6 times, with an average accuracy of 35.4%. The findings indicate that both the number of training images and the camera angle affect the performance of the helmet detection system.