Calibrating Dead Reckoning with Deep Reinforcement Learning

In general, as the moving distance increases, positioning accuracy degradation due to cumulative error is pointed out as a major problem in DR (Dead Reckoning). The distancebased positioning system has the advantage of high positioning accuracy depending on the sensor used. However, in environments where the use of infrastructure is restricted, such as security facilities or disaster environments, its utilization is limited, and Line of Sight (LOS) between devices must be guaranteed. To compensate for those limitations, we propose a framework for resetting the initial point of an object in DR. In order to reduce the position error that increases in proportion to the accumulated distance of DR, a method of resetting the starting point of DR to the position derived by distance-based trilateration was adopted. However, the process of resetting every moment when the trilateration-based position is derived can increase the system overhead and cause a positioning delay. Therefore, we derived the optimal starting point reset position and frequency by using Deep Reinforcement Learning (DRL). Through simulation, we verified that the proposed system improves the positioning accuracy compared to conventional DR through low resource consumption.