Wearable Computing Approach for Indoor Positioning

Existing approaches for indoor navigation using wearable sensors weigh heavily on path to map matching techniques. This work is an attempt to identify the user's location within a pre-selected portion of a building from users' own activities. The study employs the concept of activity-mapping for locating the user. A comparative study of various machine learning classifiers for human activity recognition was conducted and found the Multiclass Forest Classifier as the most suitable classifier for this study. A supervised dataset for training the model was collected using a bespoke Android companion application running on Samsung Galaxy Core Prime mobile and Samsung Gear Live smartwatch. This data consisted of 20 samples each of 15 user activities, totaling 15346 sample points of accelerometer readings. The trained model has an average accuracy of 0.936209 and overall accuracy of 0.521569. The decrease in overall accuracy is due to the similarity of certain activities such as cabinet-open-close and fridge-open-close. Accuracies of each observed user activities are left-turn-walk: 34.4%, right-turn-walk: 28.7%, going-down-stairs: 59.9%, going-up-stairs: 44.2%, elevator-up:50.2 %, elevator-down: 63.1%, elevator-button-press: 49.7%, cabinet-open-close: 64.8%, standing: 37.1%, U-turn-walk: 35.8%, fridge-open-close: 52.0%, walking: 36.1%, switch-on-off: 45.4%, waving: 65.6% and open-close-door: 56.3%. A portion of a building for indoor positioning was implemented as a graph data structure for computation. The implementation was tested and evaluated by conducting 22 indoor positioning experiments consisting 7793 accelerometer readings. It is observed that the accuracy of human activity recognition model directly influences the accuracy of the indoor positioning.