Pedestrian Heading Estimation Based on F-LSTM Neural Network

When global navigation satellite system navigation signals are restricted or unavailable, using the inertial measurement unit and magnetometer of a smartphone for pedestrian dead reckoning (PDR) is a simple and highly practical navigation positioning solution. Since heading estimation is the main source of error in PDR positioning, this paper proposes a correction scheme based on the Long Short Term Memory (LSTM) neural network to solve the problem of heading angle estimation errors when using a smartphone for PDR. The proposed F-LSTM (Fully Connected LSTM) network is trained on data collected in various heading directions to regressively approximate the non-linear sensor error. The data collection is relatively small, and the training time is short, which improves the practicality of the model. Experimental results show that the proposed algorithm can effectively improve the accuracy of heading angle estimation, with the heading estimation error less than 4°. The algorithm is integrated with classical PDR. The positioning error that is less than 2.44 % of the total distance.