A Deep-Learning-Based Monocular VO/PDR Integrated Indoor Localization Algorithm Using Smartphone

Visual-inertial odometer (VIO) can enable localization and navigation in indoor environments without the support of infrastructure, however the realization of VIO on smartphones is still a challenge. To build a highly precise smartphone-based VIO for pedestrian localization, a deep learning (DL)-based integrated localization algorithm is proposed, which uses the measurements of visual odometer (VO) and pedestrian dead reckoning (PDR). Specifically, the IMU data is used to estimate pedestrian’s location based on PDR, which can be readily implemented on smartphone. To address the scale ambiguity of monocular VO, this article innovatively proposes to use the gray wolf optimization (GWO) algorithm to determine the scale factor. Two effective DL algorithms, namely, back propagation (BP) and long short-term memory (LSTM) neural networks, are utilized to integrate the position estimate of monocular VO and that of PDR, and the proposed localization algorithm is named VP-BGL for convenience. Experimental data were collected from three different indoor scenarios for testing our proposed VP-BGL algorithm, and the three indoor scenarios were in a classroom, in an underground parking lot, and on a floor of an office building. The experimental results show that our proposed VP-BGL integrated localization algorithm has the best positioning compared to three exist VIOs. Compared to the state of art VIO, our proposed algorithm improves the accuracy in terms of root mean square error (RMSE) by 0.424 m on average for the three experimental fields.