Activity Recognition from Accelerometer Data

Activity recognition is a particularly salient problem, both in terms of its many practical applications and the research issues it raises. Practical applications include, for example, post-operative monitoring in ambulatory medicine, detection of health-related anomalies, prevention of falls or, for lighter purposes, monitoring of daily physical activity via an estimate of caloric expenditure or tracking of sporting activities (duration of a run, distance covered, etc.).

There are several solutions for capturing a person's activity using sensors. One possibility is to use images or videos to recognize activity, but using visual data has two major drawbacks: firstly, the person must remain in the camera's field of view, and secondly, some users may be reluctant to be constantly under the eye of a camera in their private life. Another possibility is therefore to use sensors worn by the user. In this respect, because of its democratization, discretion and computing power, the smartphone can be used to collect data on individuals, and this data can be used for activity recognition. Most smartphones contain numerous sensors, such as accelerometers, gyroscopes, magnetometers and so on. In addition to smartphones, devices can be developed specifically for activity recognition. There are, for example, wearable devices that directly integrate an accelerometer, gyroscope and magnetometer. In other cases, in addition to the various measurements derived from the above sensors, physiological signals such as heart rate, respiratory rate and oxygen saturation can also be used. In this article, we assume that we are only considering data provided by an accelerometer. The accelerometer is the most widely used wearable sensor for activity recognition. The use of this sensor means that data can be collected from any location, and is no longer dependent on the user's environment.

In this article, we focus on the procedure for recognizing activities based on accelerometric data. This procedure consists of a sequence of different stages, the final result of which is conditioned by the user's choices at each stage. The aim of this article is to provide an overview of this procedure, detailing the technological elements of each stage and the usual choices made in activity recognition.