Geomagnetic Field Based Indoor Localization Using Recurrent Neural Networks

The time-varying, unstable nature of RF signals has limited the accuracy of RF-based indoor positioning techniques such as Wi-Fi fingerprinting. Positioning errors of over 10 meters are reported in large-scale indoor environment such as airport and department stores. Compared to RF or ultrasound signals, the geomagnetic field signal exhibits stable signal strength in time domain. However, the existing geomagnetic field based indoor localization still relies on the fingerprinting technique, which is borrowed from the RF-based indoor positioning. This cannot resolve the distribution of the same geomagnetic field values and thus became a major reason for diminished performance of geomagnetic- based indoor localization. In this paper, we propose a novel indoor localization technique that uses magnetometer sensor readings as input to the artificial neural network models. The idea is that although there can be multiple locations having the same magnetic field value, as a pedestrian moves the sequence of magnetic field values will lead to a unique pattern of the sensor readings over time. We use a recurrent neural network (RNN) since it can characterize a particular location based on the current input as well as the past sequence of inputs. We first build a geomagnetic field map on our campus test-bed. Then, we generate a million traces of various pedestrian walking patterns from the map. We use Google Tensorflow with NVIDIA cuDNN library as a Deep Learning framework. 95% of the traces are used for training and 5% of the traces are used for localization evaluation. In this preliminary evaluation, we show an average positioning error of 1.062 meters compared to the average error of 3.14 meters of our BLE fingerprinting results. We cannot only improve the localization accuracy but we are also able to address the problem of continuous route tracking, which was not possible with the RF-based fingerprinting.