Rhythm Modelling of Long-Term Activity Data

by Dr. Marko Borazio


Referees: Prof. Dr. Kristof Van Laerhoven and Prof. Dr. Antonio Krüger





URI: http://tuprints.ulb.tu-darmstadt.de/id/eprint/4290



Long-term monitoring for activity recognition opens up new possibilities for deriving characteristics from the data, such as daily activity rhythms and certain quality measures for the activity performed or for identifying similarities or differences in daily routines. This thesis investigates the detection of activities with wearable sensors and addresses two major challenges in particular: The modelling of a person’s behaviour into rhythmic patterns and the detection of high-level activities, e.g., having lunch or sleeping. To meet these challenges, this thesis makes the following contributions:


First, we study different platforms that are suitable for long-term data recording: A wrist-worn sensor and mobile phones. The latter has shown different carrying behaviours for various users. This has to be considered in ubiquitous systems for accurately recognizing the user’s context. We evaluate our findings in a study with a wrist-worn accelerometer by correlating with the inertial data of a smart phone. 


Second, we investigate datasets that exhibit rhythmic patterns to be used for recognizing high-level activities. Such statistical information obtained over a population is collected with time use surveys which describe how often certain activities are performed

 by the user. From such datasets we extract features like time and location to describe which activities are detectable by making use of prior information, showing also the benefits and limits of such data.


Third, in order to improve on the recognition rates of high-level activities from wearable sensor data only, we propose the use of the aforementioned prior information from time use data. In our approach we investigate the results of a common classifier for several high-level activities, after which we compare them to the outcome of a maximum-likelihood estimation on the time use survey data. In a last step, we show how these two classification approaches are fused to raise the recognition rates.


In a fourth contribution we introduce a recording platform to capture sleep and sleep behaviour in the user’s common environment, enabling the unobtrusive monitoring of patterns over several weeks. We use a wrist-worn sensor to record inertial data from which we extract sleep segments. For this purpose, we present three different sleep detection approaches: A Gaussian-, generative model- and stationary segments-based algorithm are evaluated and are found to exhibit different accuracies for detecting sleep. The latter algorithm is pitted against two clinically evaluated sleep detection approaches, indicating that we are able to reach an optimum trade-off between sleep and wake segments, while the two common algorithms tend to overestimate sleep. Further, we investigate the rhythmic patterns within sleep: We classify sleep postures and detect muscle contractions with a high confidence, enabling physicians to efficiently browse through the data.