Human activity recognition using smartphone accelerometer and machine learning

Abstract

The stunning development in cellphone has brought a major way for extracting information. Cellphones contains many sensors with huge capabilities for extracting significant data. As a result, one can mine sensor data to recognize users activities without any additional input. Several probability-based algorithms have been used to build activity models. The idea of this thesis is to analyze the information provided by an accelerometer, and compare some machine learning algorithms to identify what the user is performing. To implement this system, we made an experiment on 30 volunteers using an Android cellphone in the right trouser pocket while performing 5 activities: Standing, Sitting, Laying, Walking and Jumping for 120 seconds. The data was also passed through a Butterworth low pass filter to eliminate noise then 21 features were extracted to analyze this data. Finally, we evaluated several classification algorithms from the area of Machine Learning including Decision Trees, K-Nearest Neighbor, Linear Discriminant and Support Vector Machine to compare the different outputs of these different rules and found that weighted K-Nearest Neighbor is the most accurate method with 99.7% of success rate. This work is useful since it introduces the steps for human activity recognition and it is a one step ahead for making many helpful applications such as fitness monitoring and security password recognition uses. Finally, we have highlighted some future work and research directions for human activity recognition using a smartphone.