A semi-supervised deep learning approach towards localization of crowdsourced data

Abstract

The burden of fingerprints building and calibration is the main bottleneck in complex indoor environments. Crowdsourcing is an attractive strategy to alleviate the burden. However, most existing approaches of estimating crowdsourced samples' locations are based on shallow machine learning architecture, which cannot fully characterize the relationship between RSS samples and locations. Consider the ability of deep learning structure in extracting reliable features from the noisy fluctuating RSS signals, we propose a transductive semi-supervised deep learning model, i.e., the unlabeled crowdsourced data can aid the training process of the model, and then the trained model can further improve the localization accuracy of the unlabeled data. First, a labeled crowdsourced sample augmentation technique is proposed to overcome the overfitting of artificial neural network (ANN) caused by insufficient data. Then, augmented labeled samples and unlabeled crowdsourced samples are utilized to train the semi-supervised ANN to construct a robust estimator. Finally, unlabeled crowdsourced samples are input into the estimator to obtain estimated labels. Experimental results show that the proposed method effectively improves the accuracy of floor detection compared with existing conventional approaches using crowdsoucred unlabeled samples.