End-user modification and correction of home activity recognition

Abstract

Sensor-enabled computer systems capable of recognizing specific activities taking place in the home may enable a host of "context-aware" applications such as health monitoring, home automation, remote presence, and on-demand information and learning, among others. Current state-of-the-art systems can achieve close to 90% accuracy in certain situations, but the decision processes involved in this recognition are too complex for the end-users of the home to understand. Even at 90% accuracy, errors are inevitable and frequent, and when they do occur the end-users have no tools to understand the cause of errors or to correct them. Instead of such complex approaches, this work proposes and evaluates a simplified, user-centric activity recognition system that can be understood, modified, and improved by the occupants of a context-aware home. The system, named Distinguish, relies on high-level, common sense information to construct activity models used in recognition. These models are transferable between homes and can be modified on a mobile phone-sized screen. Observations are reported from a pilot evaluation of Distinguish on naturalistic data gathered continuously from an instrumented home over a period of a month. Without any knowledge of the target home or its occupant's behaviors and no training data other than common sense information contributed by web users, the system achieved a baseline activity recognition accuracy of 20% with 51 target activities. A user test with 10 participants demonstrated that end-users were able to not only understand the cause of the errors, but with a few minutes of effort were also able to improve the system's accuracy in recognizing a particular activity from 12.5% to 52.3%. Based on the user study, 5 design recommendations are presented.