모바일 환경에서의 균열 탐지를 위한 효율적 모델 구성에 관한 연구

Abstract

The Pavement Management System(PMS) should be able to support proper maintenance operations for pavement distresses. In the conventional system, however, there are problems such as bottleneck operation that depends on manual processing, worsening road cracks that are left unhandled due to long-term delay of processing, and weak technological foundations for online, real-time processing. To solve this problem, this paper explores the neural network method that enables automated handling of the bottleneck process which has affected negatively on the maintenance operation of PMS. An end-to-end process that incorporates mobile environment is also proposed to cope with the evolving industry trends towards online remote detection. Since there’s no such cases in South Korea where road images collected from special cameras have been used to automate pavement detection, a study to find suitable model for mobile environment(SqueezeNet, MobileNet V3) was conducted using those images in this paper. In addition to this, to be referred by later researches on crack detection to enhance models performance, various optimization function(Adam, RAdam), activation function(ReLU, Mish), and learning rate(0.001, 0.0002) are also tested. As a result, the model trained with the combined settings for best performance is converted into TensorFlow Lite format, and made to run on the mobile environment. A prototype that shows related information for crack detection under mobile environment is also suggested. Lastly, a way to commercialize related technologies in the crack detection industry is surveyed over the result of this paper.