Smartphone motion mode recognition, using its sensors, can achieve significant improvement in many applications such as: Pedestrians navigation systems in closed areas (PDR), monitoring everyday life, security / military uses and medical uses.
In our project we used supervised ML methods in order to classify different modes: swing, talking, texting and pocket carrying while distinguish between motion upstairs and downstairs – An unexplored field of classification.In order to enable a wide and efficient use of our algorithm, we had to test the tradeoff between saving cellular battery life and model accuracy.
We found the best classifier for each sensor’s subset for maximize usage time and minimize an accuracy damage.
As sanity check we compared our results on external test set of classification on 4 states (without partition to up\down stairs) with the state-of-the-art current research and achieved better results (97%).
The classification for the 8 states mentioned above achieved accuracy of