Yong-seok Lee, Hyuk-jin Kwon, Journal of Internet Computing and Services, Vol. 25, No. 6, pp. 155-162, Dec. 2024
10.7472/jksii.2024.25.6.155, Full Text:  HTML
Keywords: AI, Multimodal, deeplearning, Computervisoin, Drone(UAV)

Abstract

This paper aims to early determine what kind of unknown track(fixed-wing drone, rotary-wing drone, balloon, flock of birds, etc.) is detected over the Korean Peninsula. The purpose of the study is to achieve success in operations by reducing reaction time and improving operational efficiency from the Joint Chiefs of Staff to battalion level echelons. To this end, the goal is to improve early reading and accuracy of unknown tracks through AI multimodal application that combines image recognition technology and data classification technology. Since the first North Korean drone crashed after performing a mission over our territory in 2014, the South Korean military's operation on the North's drone has not been successful, and in particular, accurate reading of unsymptomatic and striking weapons systems has been limited, resulting in a loss of timeliness in response and striking weapon system operation and insufficient armed response. In addition, a number of unknown targets have been identified in recent years due to the sensitive operation of surveillance sensors of military surveillance assets, but as several corps failed to read new times over the border area for few hours, they deployed and operated unnecessary personnel and equipment to increase combat fatigue, which is a major factor that hinders the operation of the best operational standby power. Therefore, To implement an AI multimodal model (image + radar data), deep learning AI technology was applied to improve the existing heuristic (visual-centered) reading process and to propose a multimodal model to build and operate the system. As a result of the proposed model experiment, 60% accuracy was obtained when using only the image recognition model, 80% accuracy when applying multimodal image and text fusion, and 90% accuracy when applying self-attention to multimodal. As shown, multimodal is an image recognition model. It was confirmed that it showed better performance.

Statistics
Cite this article