Self-Adaptation of Activity Recognition Systems to New Sensors

Embedded Intelligence, German Research Center for Artificial Intelligence, Kaiserslautern, Germany, {vitor.fortes,paul.lukowicz}@dfki.de Abstract Traditional activity recognition systems work on the basis of training, taking a fixed set of sensors into account. In this article, we focus on the question how pattern recognition can leverage new information sources without any, or with minimal user input. Thus, we present an approach for opportunistic activity recognition, where ubiquitous sensors lead to dynamically changing input spaces. Our method is a variation of well-established principles of machine learning, relying on unsupervised clustering to discover structure in data and inferring cluster labels from a small number of labeled dates in a semi-supervised manner. Elaborating the challenges, evaluations of over 3000 sensor combinations from three multiuser experiments are presented in detail and show the potential benefit of our approach. Keywords: Opportunistic Activity Recognition, Unsupervised Learning, Semi-supervised Learning, Classifier Adaptation 1. Introduction Today, state-of-the-art approaches to activity and context recognition typically assume fixed, narrowly defined system configurations dedicated to often also narrowly defined tasks. Such systems can only work when sensors are known in the training phase and they cannot adapt to new sensors in their environment. In turn, sensors are evermore present in our life, although not always available. When moving around, a person may face highly instrumented environments and places with little or no intelligent infrastructure. Concerning on-body sensing, a user may carry a varying collection of sensor enabled devices (mobile phone, watch, headset, etc.) on different, dynamically varying body locations (different pockets, wrist, bag). Thus, in order to realize their full potential, systems need to take advantage of devices that just "happen" to be in the environment, taking into account their current placement and relevance. In our previous work, we investigated how on-body position and orientation of on-body sensors can be inferred [1, 2], how position shifts can be tolerated [3], and how one sensor can replace another [4]. Preprint submitted to Computational Intelligence and Neuroscience March 15, 2018 integration. More precisely, this means to answer the question how can a new sensor's data be integrated in an existing activity recognition system at runtime in order to improve this recognition process. Extending a system that used n sensors to one that uses (n 1) has many challenges. For instance, training data is expensive, and thus we cannot expect the new (n 1) data to be labeled.