Human activity recognition using smartphone sensors: A dense neural network approach

Abstract

Smartphones are now more and more advanced and the recent smartphone generation incorporates many varied and strong sensors like linear or tri-axial, linear or tri-axial gyroscopes, GPS, etc. The accessibility of this sensor in the mass communication systems provides fascinating possibilities in information mining and machine learning; in addition, the human activity recognition a.k.a. HAR generates a fresh study arena. Wireless sensor-based HAR has an important effect on outpatient surveillance illnesses, physical activity, elderly care, sports and physical therapy. In order to detect ten human actions, we want to walk, walking upstairs, walking downstairs, sitting, standing, lying, using toilet, jogging, writing and typing, positioning the smartphone on hand. We gathered information from both the male and female subject while performing the listed tasks from labeled tri-axial accelerometer and tri-axial gyroscope. In order to achieve necessary sensor information, we developed an Android application with a frequency of 1 Hz. Then we used a model known as DNN for the Dense Neural Network to predict the operations of information from gathered sensors and gyroscope. In addition to DNN for operating studies we have also practiced Support Vector Machines, Decision Tree, K-Nearest Neighbor, Random Forest, Naïve Bayes and Logistical Regression. Our key objective was to examine a comparative analysis with statistical and profound classifiers on male and female subject. We also confirmed that information gathered at reduced frequencies can also perform well. We obtained maximum level of precision of 94.38% and 93.35% for male and female data sets only with DNN model respectively.