Implicit neural representation based SLAM with semantic information

Abstract

Simultaneous Localization and Mapping(SLAM) field, which simultaneously estimates camera pose and mapping, has been one of the main research fields in autonomous driving, intelligent robots, and extended reality today. Recently, slam studies of implicit representation methods have been actively researched to improve huge memory usage problem from explicit representation based SLAM methods. However, in implicit representation SLAM system, the estimation of camera accuracy is relatively low compared to existing explicit representation based slam studies. Meanwhile, as simply estimating the location and mapping of the camera is not enough for an immersive extended reality experience, the need for digital twin with semantic information for individual objects emerges. This leads to more semantic information, which could be used in SLAM system, is acquired in advance. Therefore, in this paper, assuming that there is an accurate semantic information acquired in advance, we propose a system to improve the insufficient camera accuracy of implicit neural representation based slam system.