Exploring unconstrained mobile sensor based human activity recognition

Human activity recognition using data from wearable sensors, and more recently, mobile and smartphone sensors, is a widely researched problem. In this study, we explore the detection of human activities (walk, run, jog, climb stairs up, climb stairs down, stand still) from data acquired using smartphone sensors, accelerometer, gyroscope and magnetometer. We address the task of detecting activities irrespective of the user specific mobile sensor position, orientation and body attachment. We compare the classification accuracies in case of orientation corrected and uncorrected data in different positions and orientations across multiple individuals. Also, we show that a single device and accelerometer data is sufficient to detect the activities. Additionally we compare activity detection accuracies when the user holds the device in the hand or carries the smartphone in the pockets, as in a real life scenario, by synchronously recording data from three mobile sensors devices. Finally we discuss the results and inferences of our experiments.