A long-term evaluation of sensing modalities for activity recognition

Abstract

We study activity recognition using 104 hours of annotated data collected from a person living in an instrumented home. The home contained over 900 sensor inputs, including wired reed switches, current and water flow inputs, object and person motion detectors, and RFID tags. Our aim was to compare different sensor modalities on data that approached "real world" conditions, where the subject and annotator were unaffiliated with the authors. We found that 10 infra-red motion detectors outperformed the other sensors on many of the activities studied, especially those that were typically performed in the same location. However, several activities, in particular "eating" and "reading" were difficult to detect, and we lacked data to study many fine-grained activities. We characterize a number of issues important for designing activity detection systems that may not have been as evident in prior work when data was collected under more controlled conditions. © Springer-Verlag Berlin Heidelberg 2007.