Detecting Physical Activities Using Body-Worn Accelerometers

This paper addresses the task of using data from accelerometers attached to a person’s body to determine the kind of physical activity being performed by that person. The activities of interest are routine ones such as sitting, walking up a flight of stairs, walking, and jogging. The paper describes methods for segmenting the time-series data from accelerometers and for extracting features that are effective for determining activities when used in conjunction with well established classification algorithms. These methods are implemented in a prototype that is used to evaluate their effectiveness on a publicly available dataset of tagged accelerometer traces. The prototype also provides intuitive visualizations of the accelerometer traces, allowing a human expert to gain a better understanding of both the dataset and the predictions from the classifiers. Although the methods in this paper use fewer and simpler features extracted from the raw accelerometer data, they provide higher accuracies when compared to those reported in prior work on the experimental dataset.