A Feasible Model Training for LSTM-Based Dual Foot-Mounted Pedestrian INS

Abstract

Deep learning (DL) has been confirmed as an effective method to develop inertial measurement unit (IMU) based pedestrian inertial navigation system (INS). Nevertheless, collecting data for training the DL models is always a challenge. Conventional motion capture systems are expensive and they can be applicable within a restricted range. The real time kinematic-global positioning system (RTK-GPS) has concerns of low data collection rate and outdoor usage limitations. Hence, this paper presents a feasible and easily deployable hand-push odometer platform (HPOP) that was modified from a conventional wheeled walker. The 30Hz HPOP speed information is arranged by combining the dual foot-mounted IMUs’ data for the training of long short-term memory (LSTM) models to develop a pedestrian walking speed estimator, where the training dataset contains 858,751 data items. Moreover, the Fick angle is further utilized with the estimated walking speed to form a pedestrian INS. In a 2m*2.6m rectangle path, the absolute path tracking error was 0.1024m; the RMSE of walking speed was 0.04768m/s; path walking distance error was 0.089m. In a 52.46m*8.16m basement corridor area, a 1.06m homing positioning error was investigated in a 136.6m round trip corridor path experiment.