Human Activity Recognition with Wearable Sensors

• Main Author: Huynh, Duy Tâm Gilles
• Format: Others
• Language: German; en
• Published: 2008
• Online Access: https://tuprints.ulb.tu-darmstadt.de/1132/1/Dissertation_Tam_Huynh.pdf; Huynh, Duy Tâm Gilles <http://tuprints.ulb.tu-darmstadt.de/view/person/Huynh=3ADuy_T=E2m_Gilles=3A=3A.html> (2008): Human Activity Recognition with Wearable Sensors.Darmstadt, Technische Universität, [Ph.D. Thesis]

This thesis investigates the use of wearable sensors to recognize human activity. The activity of the user is one example of context information -- others include the user's location or the state of his environment -- which can help computer applications to adapt to the user depending on the situation. In this thesis we use wearable sensors -- mainly accelerometers -- to record, model and recognize human activities. Using wearable sensors allows continuous recording of activities across different locations and independent from external infrastructure. There are many possible applications for activity recognition with wearable sensors, for instance in the areas of healthcare, elderly care, personal fitness, entertainment, or performing arts. In this thesis we focus on two particular research challenges in activity recognition, namely the need for less supervision, and the recognition of high-level activities. We make several contributions towards addressing these challenges. Our first contribution is an analysis of features for activity recognition. Using a data set of activities such as walking, standing, sitting, or hopping, we analyze the performance of commonly used features and window lengths over which the features are computed. Our results indicate that different features perform well for different activities, and that in order to achieve best recognition performance, features and window lengths should be chosen specific for each activity. In order to reduce the need for labeled training data, we propose an unsupervised algorithm which can discover structure in unlabeled recordings of activities. The approach identifies correlated subsets in feature space, and represents these subsets with low-dimensional models. We show that the discovered subsets often correspond to distinct activities, and that the resulting models can be used for recognition of activities in unknown data. In a separate study, we show that the approach can be effectively deployed in a semi-supervised learning framework. More specifically, we combine the approach with a discriminant classifier, and show that this scheme allows high recognition rates even when using only a small amount of labeled training data. Recognition of higher-level activities such as shopping, doing housework, or commuting is challenging, as these activities are composed of changing sub-activities and vary strongly across individuals. We present one study in which we recorded 10h of three different high-level activities, investigating to which extent methods for low-level activities can be scaled to the recognition of high-level activities. Our results indicate that for settings as ours, traditional supervised approaches in combination with data from wearable accelerometers can achieve recognition rates of more than 90%. While unsupervised techniques are desirable for short-term activities, they become crucial for long-term activities, for which annotation is often impractical or impossible. To this end we propose an unsupervised approach based on topic models that allows to discover high-level structure in human activity data. The discovered activity patterns correlate with daily routines such as commuting, office work, or lunch routine, and they can be used to recognize such routines in unknown data.