Mathias Stäger

Recognizing Workshop Activity Using Body Worn Microphones and Accelerometers

The paper presents a technique to automatically track the progress of maintenance or assembly tasks using body worn sensors. The technique is based on a novel way of combining data from accelerometers with simple frequency matching sound classification. This includes the intensity analysis of signals from microphones at different body locations to correlate environmental sounds with user activity. To evaluate our method we apply it to activities in a wood shop. On a simulated assembly task our system can successfully segment and identify most shop activities in a continuous data stream with zero false positives and 84.4% accuracy.

Analysis of chewing sounds for dietary monitoring

The paper reports the results of the first stage of our work on an automatic dietary monitoring system. The work is part of a large European project on using ubiquitous systems to support healthy lifestyle and cardiovascular disease prevention. We demonstrate that sound from the users mouth can be used to detect that he/she is eating. The paper also shows how different kinds of food can be recognized by analyzing chewing sounds. The sounds are acquired with a microphone located inside the ear canal. This is an unobtrusive location widely accepted in other applications (hearing aids, headsets). To validate our method we present experimental results containing 3500 seconds of chewing data from four subjects on four different food types typically found in a meal. Up to 99% accuracy is achieved on eating recognition and between 80% to 100% on food type classification.

WearNET: A Distributed Multi-sensor System for Context Aware Wearables

This paper describes a distributed, multi-sensor system architecture designed to provide a wearable computer with a wide range of complex context information. Starting from an analysis of useful high level context information we present a top down design that focuses on the peculiarities of wearable applications. Thus, our design devotes particular attention to sensor placement, system partitioning as well as resource requirements given by the power consumption, computational intensity and communication overhead. We describe an implementation of our architecture and initial experimental results obtained with the system.

Implementation and evaluation of a low-power sound-based user activity recognition system

The paper presents a prototype of a wearable, sound-analysis based, user activity recognition device. It focuses on low-power realization suitable for a miniaturized implementation. We describe a tradeoff analysis between recognition performance and computation complexity. Furthermore, we present the hardware prototype and the experimental evaluation of its recognition performance. This includes frame by frame recognition, event detection in a continuous data stream and the influence of background noise.

Titan: A Tiny Task Network for Dynamically Reconfigurable Heterogeneous Sensor Networks

Context recognition, such as gesture or activity recognition, is a key mechanism that enables ubiquitous computing systems to proactively support users. It becomes challenging in unconstrained environments such as those encountered in daily living, where it has to deal with heterogeneous networks, changing sensor availability, communication capabilities, and available processing power.

Power and accuracy trade-offs in sound-based context recognition systems

This paper presents an empirical design methodology to optimize a context recognition system with respect to a trade-off between power consumption and recognition performance rather than straightforward maximization of the recognition rate. As illustration, we present a case study in which the interaction with different household appliances is detected by means of a wrist worn microphone and accelerometers. This example, which is embedded in the larger context of an assisted living scenario, demonstrates that the proposed method leads to improvements in battery lifetime by a factor of 2-4 with only little degradation in recognition performance. For a specific sensor node, we show that a recognition rate of 94% can be achieved with a power consumption of just 3.3 mW, resulting in a battery lifetime of 168 h.

SoundButton: design of a low power wearable audio classification system

The paper deals with the design of a sound recognitionsystem focused on an ultra low power hardware implementationin a button like miniature form factor. We present theresults of the first design phase focused on selection andexperimental evaluation of sound classes and algorithmssuitable for low power realization. We also present theVHDL model of the hardware showing that our method canbe implemented with minimal resources. Our approach isbased on spectrum analysis to distinguish between a subsetof sound sources with a clear audio signature. It alsouses intensity analysis from microphones placed at differentlocations to correlate the sounds with user activity.

Towards Wearable Autonomous Microsystems

This paper presents our work towards a wearable autonomous microsystem for context recognition. The design process needs to take into account the properties of a wearable environment in terms of sensor placement for data extraction, energy harvesting, comfort and easy integration into clothes and accessories. We suggest to encapsulate the system in an embroidery or a button. The study of a microsystem consisting of a light sensor, a microphone, an accelerometer, a microprocessor and a RF transceiver shows that it is feasible to integrate such a system in a button-like form of 12 mm diameter and 4 mm thickness. We discuss packaging and assembly aspects of such a system. Additionally, we argue that a solar cell on top of the button – together with a lithium polymer battery as energy storage – is capable to power the system even for a user who works predominantly indoors.

Dealing with Class Skew in Context Recognition

As research in context recognition moves towards more maturity and real life applications, appropriate and reliable performance metrics gain importance. This paper focuses on the issue of performance evaluation in the face of class skew (varying, unequal occurrence of individual classes), which is common for many context recognition problems. We propose to use ROC curves and Area Under the Curve (AUC) instead of the more commonly used accuracy to better account for class skew. The main contributions of the paper are to draw the attention of the community to these methods, present a theoretical analysis of their advantages for context recognition, and illustrate their performance on a real life case study.

Power and size optimized multi-sensor context recognition platform

This paper presents a miniaturized low-power platform for real-time activity recognition. The wearable sensor system comprises of accelerometers, a microphone, a light sensor and signal processing units. The recognition is performed with low-power features and a decision tree classifier. Power measurements show that our 4.15/spl times/2.75 cm/sup 2/, 9 gram platform consumes less than 3 mW and can perform continuous classification and result transmission for 129 hours on a small lithium-polymer battery.