Semi-supervised Indoor Positioning with Masked Crowdsourcing Fingerprints

The fingerprint positioning scheme has important application value and development prospects in indoor positioning. However, traditional fingerprint positioning schemes often require time-consuming and laborious fingerprint collection to improve positioning accuracy. In this study, we introduce the BERT model in Natural Language Processing (NLP) into the field of indoor positioning and propose a semi-supervised indoor positioning with masked crowdsourcing fingerprints. First, we convert the fingerprint data into text-like data to fit the model input. Then we design the enhanced BERT as a feature extractor to pretrain the masked unlabeled crowdsourcing fingerprints, namely CrowdBERT. Finally, CrowdBERT is fine-tuned with sparsely labeled fingerprints to achieve high-precision indoor positioning. Experimental results demonstrate that CrowdBERT significantly outperforms the other traditional positioning algorithms such as KNN, DNN, ResNet, SAE and VAE.