Glass Detection Network Based on RGB-Infrared Image Fusion

Abstract

Humans can easily distinguish glass in indoor and outdoor scenes. However, when robots work indoors or outdoors, they struggle to perceive the presence of glass, relying on sensors accurately. They may mistakenly interpret objects reflected by the glass as obstacles and even fail to detect the presence of glass, leading to collisions and potential hazards. Accurately identifying transparent objects like glass in the environment has become a challenging problem in robot environment perception tasks. This paper first introduces the characteristics of previous glass recognition methods. A lightweight RGB-Infrared fusion glass detection network IRGNet was designed based on low power consumption requirements and high real-time performance for mobile robots. The network includes an information fusion module (IFM) integrating complementary feature information from RGB and infrared images at multiple scales. We conducted experiments on Trans10k and the self-made dataset IRG dataset. The experimental results demonstrate that this network outperforms previous detection speed and accuracy methods.