Indoor Positioning Using Wireless Fingerprinting Based on Gaussian Models

This paper considers the application of Gaussian models for indoor positioning using wireless fingerprinting. In particular, the Gaussian kernel metric is utilized to form the cost function for selecting the nearest neighbors. Experimental data from a well-known dataset collected from a university library was applied for the study. It was found that when the standard deviation of the Gaussian kernel is increased, the locations on different floors become more separable in terms of the cost function. The estimation accuracy is significantly better when the standard deviation of the Gaussian kernel is fixed to be the same for all the access points rather than each being individually estimated from the training set. This is because the variation of the fingerprints across a longer term can be many times higher than that computed across a very short term. For this dataset, this factor ranges between 0.42 and 195, with an average of 19.78. When the number of access points is reduced, the general trends in accuracy remain unchanged. The fingerprint distance maximization method was found to outperform three other access point selection approaches. The results show that the use of the Gaussian kernel metric achieved an improvement of approximately 4% in the mean position error over the commonly used Euclidean metric.