Reliable trajectory classification using Wi-Fi signal strength in indoor scenarios

The time-series nature of human movement inside buildings can be exploited for common tasks of location-based computing. With this paper, we propose to use Wi-Fi signal strength measurements directly to infer the trajectory in comparison with a database of trajectories removing the need for accurate map information or fingerprint databases. A trajectory consists of a time-series of sensor readings of all Wi-Fi signals in reach measured by a mobile device. Starting from these measurements, we discuss several possibilities of denoising, filtering and classification of trajectories to improve our approch. By using a variant of the Douglas-Peucker algorithm we reduce the amount of computation without severe degradation of classification performance. Furthermore, we increase platform scalability by using a fast filter operation based on the Jaccard index of presence of access points to prune irrelevant trajectories early. With respect to our setting, the Fréchet-distance between trajectories has proven to be a very good choice outperforming dynamic time warping. Finally, we intorduce several data-driven trajectory segmentation schemes in order to be able to match partial trajectories early. The evaluation is based on the collection of trajectories in specific situations including staircases, hallways and movement inside a single room. With this approach, we are able to reliably classify trajectories without an intermediate step of calculating spatial position. This results in increased stability with respect to local changes in the environment, as these changes only affect a small part of a longer trajectory.