Improved Step Detection with Smartphone Handheld Mode Recognition

Abstract

Indoor positioning is becoming an essential aspect of many applications. Usually, the inertial measurement unit (IMU)-based pedestrian dead reckoning (PDR) system is preferred due to its simple and cost-effective structure. One method to enhance the accuracy of PDR is to detect the walking steps of pedestrians accurately. The error of existing step detection methods grows with changes in handheld mode. Therefore, a novel step detection method by considering smartphone handheld modes is proposed in this study. We first construct a convolutional neural network (CNN) classification model to recognize four smartphone handheld modes (texting, calling, pant's pocket, and swing). The CNN model is trained with features extracted from the IMU signals, characterized by scalogram images that are generated through the continuous wavelet transform (CWT). In the step detection phase, a threshold-based peak detection approach is used to count and detect steps. We propose a method to select the distinct signal for step detection corresponding to the smartphone handheld mode. To be specific, we use the z-axis acceleration signal in the calling, pant's pocket, and texting modes. In case of the swing mode, we use the first principal component of the x and y-axes acceleration signal, which is aligned with the horizontal direction of arm swing. After the five-fold cross validation, the results demonstrated that the CNN classifier achieved an accuracy higher than 98% in all smartphone handheld modes. Compared with a conventional method, the proposed step detection improved the accuracy around 0.14%–5.40%.