A Convolutional Neural Network Architecture for Multi-floor Indoor Localization Based on Wi-Fi Fingerprinting

Nowadays Location Based Services applications are increasingly useful. However, problems like floor identification for multi-buildings and adverse effects of devices diversity are needed to be resolved. In this paper we propose a new approach using cosine similarity computed by Wi-Fi fingerprints and radio map and using Convolutional Neural Network (CNN) model to achieve multi-floor classification. We propose in this paper to use locations-based similarity as the feature vector instead of using conventional AP sets. We also use a timesaving walk-survey method to collect Wi-Fi fingerprint. Experimental results show that our proposed CNN floor classifier has 98.37% training accuracy and 99.51% test accuracy. Compared with recent deep neural networks, our proposed approach achieves state-of-the-art floor classification accuracy but only needs a training data set almost 5 times smaller than that of other approaches.