A learning algorithm for localizing people based on wireless signal strength that uses labeled and unlabeled data

Abstract

We propose a probabilistic technique for localizing people through the signal strengths of a wireless IEEE 802.11b network. Our approach uses data labeled by ground truth position to learn a probabilistic mapping from locations to wireless signals, represented by piecewise linear Gaussian. It then uses sequences of wireless signal data (without position labels) to acquire motion models of individual people, thereby improving its ability to predict individuals' motions. As a result, the localizer becomes increasingly accurate while it is being used. The approach has been implemented in an office environment, and results are reported for a systematic study involving labeled and unlabeled data.