Comparison of Convolutional Neural Network Architecture on Detection of Helmet Use by Humans

Abstract

The helmet is one of the protective equipment for the head when driving. Although it is a protector, some criminals misuse helmets to disguise their identities, such as robbery at ATMs. Some places have put a sticker not to wear a helmet in the ATM room. However, this advice is often violated. The research adopted the Convolutional Neural Network (CNN) algorithm to identify humans who use helmets and do not use helmets based on digital images. Several CNN models, such as MobileNet-V2, ResNet-50, and VGG-16, were compared in performance. The experiment was carried out using a dataset consisting of 3,207 images which were divided into two classes. The first class is used for classifying human images using helmets with 1,603 images. At the same time, the second class is for images of humans who do not use helmets, with a total of 1,604 images. The test results show that the architecture with the highest accuracy value is ResNet-50, 97.81%. At the same time, the mobileNet-V2 architecture obtained a lower accuracy value of 96.36% and the VGG-16 architecture of 52.25%.