Activity and Location Recognition Using Wearable Sensors

C ontext awareness—determining a person’s current location and recognizing what he or she is doing—is a key functionality in many pervasive computing applications. Locationsensing techniques are based on either relative or absolute position measurements.1 Much of the current research in this area, described in the “Related Work” sidebar, uses absolute-measurement–based approaches (also called reference-based systems). However, using both relative and absolute methods, as robotics often does, is usually more effective in terms of cost and performance. The fundamental idea of the relative measurement approach is to integrate incremental motion information over time. This is known as dead reckoning or odometry. We began our project to study the feasibility of applying the dead-reckoning method to recognize a person’s location in indoor environments. We focused on detecting walking behavior, because human locomotion is achieved mainly via walking. If a system can recognize walking behaviors and count the number of steps, it can estimate a person’s current location referenced on a known starting location. As a first attempt, we suggested a combined method involving a simple active beacon and dead reckoning that could track a person’s location continuously with reasonable accuracy.2 However, it also showed an inherent problem of dead reckoning—that heading errors cause large lateralposition errors. To avoid this problem, we developed a location recognition method based not on a description of motion in 2D space but on verbal descriptions of path segments, such as “walk straight, then go down the stairway, and turn right.”3 We obtained a promising result: 86.7 percent of the average recognition ratio (the number of correctly detected transitions divided by the total number of location transitions) for 12 transitions between 6 locations in an office environment. However, the method was limited in the case of a long path, because it determined transitions based on accumulated numbers of steps instead of a whole sequence; thus, it showed the location transition before the person reached the destination. In addition, the main source of error originated from misrecognizing the person’s activity. This article suggests an improved method to tackle these limitations. This involves