Sequence-based motion recognition assisted pedestrian dead reckoning using a smartphone

Abstract

PDR (Pedestrian Dead Reckoning) is a relative positioning method using step length and heading. However, the step length is varying while a pedestrian performs different motions. With estimated motion state, the adaptive step length model can be applied. Considering that human motion is a continuous procedure, in this paper, we propose a sequence-based motion recognition method which estimates the motion states from a sequence of data. In contrast with traditional classifiers, this paper deploys a HMM (Hidden Markov Model) to infer the state labels of sequence motion. And with the assist of motion recognition, a 3D indoor pedestrian localization is presented. Experimental results show that the classification accuracy of sequence-based motion recognition is improved comparing to that of using Naïve Bayes classifier on standalone motion states. Furthermore, the positioning accuracy of a pedestrian in indoor environments is promoted using proposed method in this paper. The mean positioning error is reduced from 0.78 to 0.30 m. The 50th and 95th percentile errors are also cut down in the test within a typical office building.