Improved Indoor Positioning Using the Baum-Welch Algorithm

In this paper, we examine the exploitation of individual patterns of behavior to enhance indoor positioning of pedestrians. We make use of the fact that, due to habits and needs, a person is likely to be in some locations more often than others. For example, at their work or in their home, a person is likely to spend more time in some rooms than in others. Therefore, it seems natural to take advantage of established behavior patterns when performing indoor localization in an effort to improve accuracy. Such improvement is particularly beneficial during emergencies, where location inaccuracies may lead to life-threatening delays in response times. In this work, habitual behavior is modeled and learned using a hidden Markov model. It is shown that, applying the Markov model for location estimation results in more accurate estimates when compared to using a standard particle filter with odometry information. Additionally, transition probabilities as well as position error distributions do not need to be known a priori since they can be learned using the Baum-Welch algorithm. Results show how the Baum-Welch algorithm can even learn the distributions of biased estimates. On the other hand, it is shown how user feedback can help accelerate the learning process, while guaranteeing good parameter estimation accuracies.