Unobtrusive and ubiquitous activity recognition with enhanced contextual awareness

This thesis examines how smart devices can be used to monitor physical activities, and enhance contextual awareness using contemporary MEMS sensors. The increasing prevalence of smart devices coupled with the recent addition of a growing range of MEMS sensors raises the question of how these devices can be used in user driven preventative healthcare applications, including activity recognition. To this end, data was collected from a total of N=39 participants, over three separate trials, at three universities. A diverse range of activities were examined, including sitting, standing, walking, cycling, running, ascending and descending stairs, and elevators. A preliminary study was carried out to evaluate the effectiveness of a loosely placed smartphone, with sporadic sampling, to detect a subset of six activities. A comparison was made with a dedicated device, securely placed at the chest. A comprehensive suite of applications was developed to gather acceleration data, synchronize signals from both devices, and generate features from this data. Models were trained and evaluated using five separate classifiers, and demonstrated that both devices produced comparable results. Overall true positive classification results of 84% and 85% were achieved for smartphone and chest sensor, respectively. A subsequent feasibility study evaluated the role multi-sensor feature fusion can play in activity recognition and enhanced geospatial awareness. The contributions of up to six built-in smartphone sensors were evaluated to detect from among nine separate activities. Overall results earmarked the accelerometer as the sensor which contributed the most to the task at hand. Comparisons were made between a smartphone accelerometer, and a smartwatch accelerometer for all nine activities. Results illustrated a smartphone placed in a trouser pocket produced better results than a smartwatch placed on the wrist. In addition the study evaluated the ability of smartphone sensors to detect when participants were indoors or outdoors. Overall results were positive, with one model correctly classifying 100% of all test cases. Finally, a study to evaluate the placement of a smartphone was carried out. A better understanding of where the device is carried can improve physical activity models. Locations examined included in the hand, backpack, or pocket. Overall results of this study were as high as 96% for a balanced accelerometer dataset.