XDRNet: Deep Learning-based Pedestrian and Vehicle Dead Reckoning Using Smartphones

Abstract

As the city continues to grow, the demand for positioning in large urban buildings and underground spaces has become particularly important. However, due to the obstruction by walls and other structures, satellite signals are subject to extremely strong interference and attenuation indoors, so the indoor positioning accuracy is relatively low compared with outdoor positioning. And with the development of MEMS sensors, it is possible to use inertial navigation on mobile devices. Because of the cumulative error of the integration process, there are problems in using traditional inertial positioning methods on smartphones. In this paper, we presented a deep learning based pedestrian and vehicle Indoor positioning method, XDRNet. This method can get the positioning trajectories of moving objects by learning the relationship between real trajectories and motion states. Then, driving behavior and pedestrian behavior are distinguished by deep neural network method. In this paper, a lightweight network structure is used to make the above method work better on smartphones. Based on the above research, this paper uses the underground parking area of an office building as the experimental area to evaluated the positioning performance of above methods. The experimental results have demonstrated the superiority of the methods in this paper, which is applicable to both pedestrian and vehicle motion carriers.