Classification of Physical Activities Using a Smartphone: Evaluation Study Using Multiple Users

Abstract

Nowadays, smartphones play an ubiquitous role in accessing and processing information, most of them having a myriad of integrated sensors that makes them capable of generating information with high accuracy and precision. The monitoring of physical exercises presents itself as one of the new trends, made possible by the use of devices like smartphones. Motion sensors such as the accelerometer enable live motion measurement. This paper intends to study this issue and develop an application for the Android operating system, which takes advantage of sensors embedded in smartphones and web technologies, with the goal to classify multiple physical activities. The developed solution is based on client-server architecture. The client application performs data acquisition, visualization and recording of the signal obtained by the smartphone's accelerometer and the server application receives the information acquired by the client, processes it and classifies it. In order for the application to be able to classify multiple movements throughout the activity performed by the user, an extensive analysis of the acquired signals was carried out to understand their most distinctive features. We used a supervised approach with the goal of reviewing the best techniques that should be useful for achieving the classification with the lowest error. For the signals acquisition the smartphone was positioned along the waist, inside the right front pocket in an attempt to simulate conditions as naturally as possible. The study explored features extracted in both the time and frequency domain, and parametric and non-parametric classifiers. Preliminary results demonstrate that the classification of activities can be done with remarkable accuracy ( > 95%).

Training intensity quantification of core stability exercises based on a smartphone accelerometer

“…Then, the mean velocity of COP displacement was computed [28]. Regarding the pelvic linear acceleration, the smartphone accelerometer signal was low-pass filtered at 10 Hz (4 th -order, zero-phase-lag, Butterworth) [29] and the mean acceleration was calculated as the average of the acceleration magnitude data series [30, 31]. The computation of the COP and acceleration variables was carried out with “ad hoc” software, developed by our research group within LabView 9.0 environment (National Instruments, USA).…”

Feature Analysis to Human Activity Recognition

“…The kNN classifier generation is based on the work of Duarte et al [13]. In this research, the authors reached approximately 98% recognition rate by a 1NN classifier with the Euclidean distance metrics.…”

“…Previous studies have shown that artificial neural network (ANN), knearest neighbor (kNN) and decision tree (DT) are well applicable for HAR, see for example [5,11,13]. Therefore, to the efficiency measurement of selected features those classifiers have been applied.…”

Identification and Prediction of User Behavior Depending on the Context of the Use of Smart Mobile Devices

“…Nowadays, smartphones play an ubiquitous role in accessing and processing information, most of them having a myriad of integrated sensors that makes them capable of generating information with high accuracy and precision [28]. Collecting usage information on smartphones is difficult: The collection mechanism itself needs to be built and deployed to a diverse group of participants running a multitude of devices in the wild; data must be collected for extended periods of time to overcome novelty effects and find long-term trends.…”