Robust Smartphone Mode Recognition

Abstract

Smartphone based pedestrian dead reckoning (PDR) approach is commonly used for indoor positioning. Recognition of the smartphone mode can improve PDR positioning accuracy. In this paper, we employ machine learning classification algorithms to recognize the smartphone modes (e.g. pocket or swing) and thereby enabling the choice of a proper gain value to improve PDR positioning accuracy. In particular, we focus on two classification approaches: 1) tree based approaches: random forest, gradient boosting and CatBoost 2) neural network approaches: convolutional neural network, recurrent neural networks with long short-term memory units, gated recurrent unit and residual recurrent neural networks. Experimental results obtained using thirteen participates walking in an inhomogeneous environments and smartphone modes show successes of more than 97% in classifying the smartphone modes using neural network approaches.