Use of ECDF-based features and ensemble of classifiers to accurately detect mobility activities of people using accelerometers

We look at the problem of using accelerometer in smartphones to detect mobility activities of users. The activities are internally composed of several simple activities. One can perform the task of distinguishing the activities using classic classification techniques with two different data representations namely, statistical features and ECDF-based features. Our recommendation in this paper is to use the latter as it suits better for mobility activities. Our major contribution is to explore the challenge of class imbalance in detecting mobility activities. To handle that challenge, we propose use of an ensemble of a classification model. It improves accuracy of detection over standalone classification models. To evaluate performance of the recommended technique, we use transportation by a metro train as a running case study. We consider two activities during the metro train travel. They are (a) whether user is at a metro train station or (b) in a metro train. Our recommended technique results in precision of 98% for the case study. It is significantly more than the state-of-the-art value of 70% for a similar case study. This case study finds its applications in the area of smart city analytics, for instance, our solution could be used to estimate rush at metro stations. In the long run, it can also be used to enhance navigation services to account for delays at metro stations into their algorithms.