Automatic identification of fingerprint regions for quick and reliable location estimation

One of the drawbacks of location fingerprinting systems is the effort that is necessary to set up and update the fingerprint database. In this paper, we propose a novel approach to significantly reduce this effort. We split the area of operation into a grid of quadratic cells and then combine these cells into larger regions of similar signal properties using a clustering algorithm and a novel similarity measure. Thus, less training data is required, and it can be collected in a more efficient way: We move through the area of operation on predefined trajectories and interpolate the approximate position for each measurement. In addition, by storing only one fingerprint for each region, we reduce the computational requirements of the location fingerprinting algorithm considerably. Since the radio measurements are quite similar in such a region, it is hard to estimate the exact location within the region; thus we do not lose much accuracy by clustering. An evaluation of our approach shows that it achieves an accuracy that is sufficient for most location-based services and at the same time reduces the effort for the collection of the training data to a mere walk of the area of operation.