| Description |
Device-free localization (DFL) systems are used to locate a person in an environment by measuring the changes in received signal strength (RSS) on all the links in the network. A fingerprint-based DFL method, such as the type addressed in this thesis, collects a database of RSS fingerprints and uses a machine learning classifier to determine a person's location. However, as the environment changes over time due to furniture or other objects being moved, the RSS fingerprints diverge further and further from those stored in the database, causing the accuracy of the system to suffer. This thesis investigates the degradation over time of localization accuracy in radio frequency (RF) sensor networks using a fingerprint-based method with a machine learning classifier. We perform extensive experiments that allow quantification of how changes in an environment affect accuracy, through a process of moving specific items in a residential home one at a time and conducting separate localization experiments after each change. We find that the random forests classifier performs the best as changes are made, compared to three other classifiers tested. In addition, we present a correlation method for selecting the channel used with each link, which improves localization accuracy from an average of 95.2% to an overall 98.4% accuracy using random forests. We thus demonstrate that combining the random forests classifier with a correlation method of selecting a channel for each link offers a viable approach to developing a more robust system for device-free localization that is less susceptible to changes in the environment. |
