Description of Invention
Fingerprint-based indoor positioning systems (IPS) is a promising solution for GPS-denied environments. The existing method employs convolutional neural network-based (CNN) classifier to achieve accurate localization. Nevertheless, CNN-based IPSs is impractical to be implemented on resource-constrained devices due to their excessively high storage and processing demands. In this work, two knowledge distilled IPSs, namely the KD-CNN-IPS and TAKD-CNN-IPS, are developed. The proposed frameworks are evaluated on multi-floor indoor environments with varying layouts. Remarkably, both proposed frameworks successfully reduce the localization error of the baseline CNN-IPS by 22.30% and 39.01%, respectively, while achieving an impressive execution time of 0.17s.