Activity recognition with smartphone support.

In this paper, the authors describe a method of accurately detecting human activity using a smartphone accelerometer paired with a dedicated chest sensor. The design, implementation, testing and validation of a custom mobility classifier are also presented. Offline analysis was carried out to compare this custom classifier to de-facto machine learning algorithms, including C4.5, CART, SVM, Multi-Layer Perceptrons, and Naïve Bayes. A series of trials were carried out in Ireland, initially involving N=6 individuals to test the feasibility of the system, before a final trial with N=24 subjects took place in the Netherlands. The protocol used and analysis of 1165min of recorded activities from these trials are described in detail in this paper. Analysis of collected data indicate that accelerometers placed in these locations, are capable of recognizing activities including sitting, standing, lying, walking, running and cycling with accuracies as high as 98%.

Multi-sensor fusion for enhanced contextual awareness of everyday activities with ubiquitous devices.

In this paper, the authors investigate the role that smart devices, including smartphones and smartwatches, can play in identifying activities of daily living. A feasibility study involving N = 10 participants was carried out to evaluate the devices' ability to differentiate between nine everyday activities. The activities examined include walking, running, cycling, standing, sitting, elevator ascents, elevator descents, stair ascents and stair descents. The authors also evaluated the ability of these devices to differentiate indoors from outdoors, with the aim of enhancing contextual awareness. Data from this study was used to train and test five well known machine learning algorithms: C4.5, CART, Naïve Bayes, Multi-Layer Perceptrons and finally Support Vector Machines. Both single and multi-sensor approaches were examined to better understand the role each sensor in the device can play in unobtrusive activity recognition. The authors found overall results to be promising, with some models correctly classifying up to 100% of all instances.

A single vs. multi-sensor approach to enhanced detection of smartphone placement.

In this paper, the authors evaluate the ability to detect on-body device placement of smartphones. A feasibility study is undertaken with N=5 participants to identify nine key locations, including in the hand, thigh and backpack, using a multitude of commonly available smartphone sensors. Sensors examined include the accelerometer, magnetometer, gyroscope, pressure and light sensors. Each sensor is examined independently, to identify the potential contributions it can offer, before a fused approach, using all sensors is adopted. A total of 139 features are generated from these sensors, and used to train five machine learning algorithms, i.e. C4.5, CART, Naïve Bayes, Multilayer Perceptrons, and Support Vector Machines. Ten-fold cross validation is used to validate these models, achieving classification results as high as 99%.