Ainandafiq Muhammad Alqadri, Chikamune Wada, Fitri Utaminingrum
Individuals with both mobility and visual impairment faces difficulty to ambulate into a room that can only be recognized through the writing of the room nameplate. Room nameplate recognition system on autonomous smart wheelchair can overcome the problem. Deep learning that offers accurate prediction capabilities, has extended to various intelligent devices. However despite its remarkable potential, deep learning poses significant risks in term of computation time. Accidents can occur if the deep learning model used in smart wheelchair has high complexity computation which can make the system inference time longer. The aims of this study is to reduce time and complexity of YOLOv8n model to detect room nameplate objects. We proposed YOLOv8n-GSM, an improved YOLOv8n model to detect room nameplate using ghost module and modules pruning to reduce model complexity. YOLOv8n-GSM successfully reduces the number of parameters by 65,36%, GFLOPs by 50%, and model size by 61,47% which resulting 28,55% shorter inference speed on NUC mini computer, and 35,71% shorter inference speed on Jetson TX2 NX, while maintain the detection accuracy on 98,9%. We hope our research will be further utilized by smart wheelchair developers. Code is available at: https://github.com/ainandafiq55/YOLOv8-GSM . Copyright © 2026. Published by Elsevier Ltd.
Brawijaya University, Veteran No. 10-11, Ketawanggede, Lowokwaru, East Java, Malang, 65145, Indonesia; Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology, Hibikino 2-4, Wakamatsu-ku, Fukuoka Prefecture, Kitakyushu, 808-0135, Japan
