A Deep Learning Approach to Detect Human Behaviour Using Smart Phones

Abstract

Smart-phones nowadays have various inbuilt sensors which include the gyroscope, accelerometer, and magnetometer. This sensor data can be used to detect the current action of the human-like sitting, walking, running, walking upstairs, walking downstairs, laying and standing. Each of the three sensors gives 3 values each - values along the x, y, and z-axes - making a total of 9 values. These values are collected at 100hz, along with GPS data. The data collected over a second can be used as the raw data from which 48 values are extracted - that will later be used as inputs for the machine learning algorithm. At first, the algorithm learns from the previously collected outputs by applying supervised learning. After the algorithm is trained and is detecting user activity with a good accuracy, an app can be used to detect if the user is sitting, standing, laying etc. while the user has the smart-phone in his/her hand. This app can be designed to optimize phone usability while driving - like enabling hands-free mode or diverting calls. Here the main focus is on human motion detection on smartphone motion sensors.