Efficient object detection using object location information for mobile augmented reality

Abstract

Recently, Mobile Augmented Reality (MAR) is highlighted as a service that is feasible with the deployment of 5G cellular system. Previous researches have focused on the efficient usage of Mobile Edge Computing (MEC) server for the MAR services or enhanced Computer Vision approaches which use Machine Learning or Deep Learning. These researches are not considering Object Detection using the object location information such as relative coordinate or absolute coordinate of objects. In this thesis, we study the feasibility of Object Detection using object location information, proposing the system which uses the MEC server to store the Object Detection results with the object location information and to fetch them for the later use based on the location information of user and object location and user interaction. Our experiments show that the proposed system shortens the Object Detection Latency about one out of ten times faster than the YOLOv3Tiny model using little CPU computation power which is a quarter of it from the Object Detection approach which uses mobile device only.