We propose an inertial sensor-based approach to activityrecognition for pedestrian indoor navigation. In the consideredscenario a mobile device is held in a hand in front of the user.The recognized activities are the ones relevant to positioning inmulti-floor buildings: walking and going up or down the stairs.To model the time dependency between consecutive activitieswe employ a Hidden Markov Model (HMM). For efficientquantization of continuous features, we apply a random forestclassifier. For verification of the proposed algorithm, we con-ducted experiments with 12 participants and 4 different mobiledevices. In our comparison to state-of-the-art approaches, weimplement and evaluate major classification algorithms, suchas nearest neighbour, decision tree and dynamic BayesianNetwork. In the experiments we show the trade-off betweencomputational complexity and classification performance. Fur-thermore, we demonstrate that the complexity of the HMMcan be significantly reduced by replacing it with a dynamicBayesian network with negligible impact on classificationperformance. The best of our proposed classifier achieves aclassification accuracy of 91% for new users, which offers a30% improvement compared to state-of-the-art approaches.
For all the details, please take a look at the full paper.