Human Activity Recognition Models and Step Counter With Smartphone Sensor Data

Abstract: In this time of technology, with the availability of wearable sensors and copiousamounts of cheap data, new uses of machine learning emerge. Tasks that were previouslyheld-back by a lack of data and computation power are today more feasible and useful thanever. Human activity recognition (HAR) is one such task. HAR technology is especiallysought after in the fitness industry, and in health- and elderly care where monitoring physicalactivity is of importance. Artificial intelligence's ability to learn complex patterns makes it arequisite tool in HAR and recognizing activities from sensor data. In this project, wedeveloped and implemented a step counting algorithm as well as two machine learningmodels that can classify simple activities such as walking, running and climbing stairs. Themodels as well as the step counter use data from smartphone accelerometers andgyroscopes. Data was collected by the two participants and was preprocessed before beingused to build the algorithm and models. The step counter achieved an overall accuracy of91.7% when tested on different activities, signal lengths and positions of the smartphone.The HAR models were implemented with the Random Forest and Gradient Boostingmethods and obtained a test accuracy of 98.3% and 97.7%, respectively.