Transfer Learning for Convolutional Indoor Positioning Systems

Abstract

Fingerprinting is a widely used technique in indoor positioning, mainly due to its simplicity. Usually, this technique is used with the deterministic k - Nearest Neighbors (k-NN) algorithm. Utilizing a neural network model for fingerprinting positioning purposes can greatly improve the prediction speed compared to the k-NN approach, but requires a voluminous training dataset to achieve comparable performance. In many indoor positioning datasets, the number of samples is only at a level of hundreds, which results in poor performance of the neural network solution. In this work, we develop a novel algorithm based on a transfer learning approach, which combines samples from 15 different Wi-Fi RSS indoor positioning datasets, to train a single convolutional neural network model, which learns the common patterns in the combined data. The proposed model is then fine-tuned to optimally fit the individual databases. We show that the proposed solution reduces the positioning error by up to 25% compared to the benchmark model while reducing the number of outlier predictions.