Improving Indoor BlueTooth Localization By Using Bayesian Reasoning To Explore System Parameters
With the advent of smaller, more mobile electronic devices, a wide variety of services can now be augmented with the additional context that is provided by positional information. Systems commonly used for outdoor localization, such as GPS, cannot necessarily be used for indoor localization because often, separating a localizing device from system infrastructure with walls and other obstacles lowers accuracy. Instead, indoor localization systems can be deployed to replace the contextual information required for some situated services, that would otherwise be lost when a device moves indoors. For example, the trilateration algorithm that GPS uses to combine distance estimates from satellites can be repeated using Bluetooth (BT) devices spread throughout an environment. The signal strength of a set of beacons can be read by a localizing device, and those signal strengths can be equated to the distance between the localizing device and the beacon. These distances can then be combined using trilateration. A major source of error in such a system is that BT signal strength does not map directly to only one distance. Because microwave frequency propagation is susceptible to multipath effects and antenna direction, two devices at a fixed location can read a variety of signal strengths, which may not map to the ideal line-of-sight calibrated value. Therefore, any given signal strength reading cannot be interpreted as a single distance without introducing the potential for substantial error. One solution is to probabilistically model the relationship between distance and signal strength by modelling BT localization using a Bayesian network. In a Bayesian network, the distance versus signal strength relationship is stored as the conditional probability of a signal strength reading given a specific distance. Using a Bayesian inference algorithm, one can then reason backwards from a signal strength to a probability distribution representing the estimated position of the localizing BT device. In this thesis, I explore some of the effects of modelling BT localization with a Bayesian network. I first extend the probabilistic calibration to include the influence of the relative orientation of device antennae on the attenuation of BT signal strength between them. I then experiment with the effects of the position of a receiver within a discrete spatial bin, and of the proximity of the transmitters to the edges of the discrete space, because both have the potential to reduce the accuracy of localization using discrete variables. I found that neither affected the localization results in a significant, avoidable fashion. I then studied the effects of the scope of calibration, in terms of the number of distance values used, and of the number of beacons used in localization. I found that additional distance values and a smaller minimum distance used in calibration could result in increased BT localization accuracy, whereas many BT localization systems perform little calibration at distances smaller than 2 m. I also found that accuracy increased when the number of beacons was greater than four, and that accuracy did not significantly decrease when the number of beacons was three or fewer; whereas most trilateration systems use only three or four beacons. I conclude in general that a combination of probabilistic trilateration calibration and Bayesian network inference are viable techniques, and could allow for improvements to localization accuracy in a number of areas.
DegreeMaster of Science (M.Sc.)
SupervisorHorsch, Michael C.; Stanley, Kevin G.
CommitteeEager, Derek; Bell, Scott; Kamal, Mohammad
Copyright DateNovember 2012