2021 IEEE 32nd Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC) | 2021
Low-Effort Deep Learning Method Trained through Virtual Trajectories for Indoor Tracking
Abstract
We develop a novel low-effort Wi-Fi fingerprinting method to generate a fully labeled dataset of received signal strength indicator (RSSI) values to train a deep learning model for indoor localization and trajectory estimation. The positioning accuracy of Wi-Fi fingerprinting approach can be improved by collecting large amount of labeled data with diverse set of hardware devices, orientation angles, and environmental conditions. However, the cost of data collection becomes prohibitive. In crowdsourcing method, volunteers who frequently visit the indoor space, contribute unlabeled but diverse trajectories. However, despite the potential to collect large quantities of data, unlabeled trajectories produce large positioning errors. In this paper, we propose a novel method where we only collect a base-fingerprinting set with a limited number of devices, orientation angles, and environmental conditions. Then, we develop RSSI model which allows us to simulate device & orientation angle and environmental conditions diversity as noises that are added on top of RSSI values in the base-fingerprinting set. We then generate a very large set of fully labeled diverse virtual trajectories to train appropriate deep learning models for use in the online-phase of Wi-Fi fingerprinting method. In this paper, we train 1-D convolutional neural network (1-D CNN) models called base- and diverse-models. Base-model is trained on the virtual trajectories obtained exclusively from the base-fingerprinting set. On the other hand, diverse-model is trained on the so-called noisy trajectories which are created with the help of RSSI model. To validate the effectiveness of our approach, we perform experiments in our campus library. When the online-phase trajectory is coming from hardware & orientation angles and environmental conditions not used for data collection, the diverse-model achieves an average mean square error of 1.24m as compared to 19.25m for the base-model. These results demonstrate the effectiveness and simplicity of our proposed approach.