Indoor Navigation
This work was motivated by the goal of building a hand-held navigation system that could guide people around urban environments, similar to the car-based navigation systems that are popular today. We have focused on the task of inferring a person's location even in situations in which Global Positioning Systems (GPS) cannot provide this information, such as when the person is indoors or in crowded urban areas where there is no line of site to the GPS satellites. As an alternative to installing active badges or beacon systems, we have developed a system which navigates based on naturally occurring landmarks, such as magnetic fields from steel beams in walls, fixed arrangements of fluorescent lights, and temperature gradients across rooms.
Background & Objective: This work was initiated by discussions about Intelligent Transportation Systems (ITS) at Sanken in 1998. The overall objective is to extend car navigation systems so that they can be removed from the car and carried around by the user.
Technical Discussion: In general, we are interested in applying machine-learning techniques to the task of inferring aspects of the user's state given a stream of inputs from sensors worn or carried by a person. We are especially interested in integrating information from the diverse set of cheap, lightweight sensors that are now available, including accelerometers, temperature sensors, and photoresistors. We found that the raw sensor signals were unsuitable for use as direct inputs to a machine-learning algorithm. Our navigation algorithm performs very poorly, with almost 50% error, if we use only the raw sensor signals. The reason is that there is too great a distance between the low-level raw signals and the high-level inference we wish to make. To address this, we introduce a "data cooking" module that computes appropriate high-level features from the raw sensor data. These high-level features do not add any new knowledge to the system; they simply reformulate the existing information into a form in which it can be used more effectively by the machine-learning algorithm. Introducing these high-level features was found to improve performance on the indoor-navigation task dramatically. By introducing these high-level features, we are able to reduce the error rate to 2% in our example environment.
Technology Area: Artificial Intelligence
Modification Date: January 23, 2007