Feature Optimization Integrated With Hybrid Regression Based Machine Learning Using Received Signal Strength Measurements for Indoor Localization

In this article, a new indoor positioning algorithm with received signal strength indicators (RSSI) fingerprints by feature optimization and hybrid regression (HR) model is proposed. In the offline phase, the gaussian filtering and normalization are utilized to pre-processing training set. Immediately after, the unsupervised clustering Kmeans ++ algorithm is applied to obtain the labels of each measurement and the measurement label set is used for classification learning by the support vector classification(SVC) method. Moreover, ReliefF(RF) and Principal Component Analysis (PCA) based feature optimization is applied to extract features from each measurement-position. Finally, HR learning is performed. In the online phase, the fingerprint scoring program (FSP) is used to assess the training set. Then, the coordinates of test points are predicted by RFPCA-HR algorithm. The field tests show that, with reducing the volume of data by about 98%, the positioning system has an average positioning error of 0.92m and the maximum positioning error is controlled within 2m. This method has a certain improvement compared with the previous methods.