Learning Indoor Area Localization: The Trade-Off Between Expressiveness and Reliability

Abstract

Fingerprinting-based algorithms are a cost-effective approach to obtain coarse location information of a person within an indoor space. Fingerprints consist of observable sensor signals that emerge from the existing sensing infrastructure within a building and differ depending on the location within the building. The relation between a fingerprint and its location can be learned in a supervised setting, such that the learned model is able to estimate a location for an unseen fingerprint. The most prominent technology that composes the fingerprint is WLAN, where the observed feature is commonly the received signal strength (RSS) to all observable access points. Since RSS suffers from signal defects such as multi-path propagation, learning a model that exactly pinpoints the location is hardly possible. To still achieve a reliable prediction, models can be specifically trained to provide a coarser location estimate (e.g., area/zone) with a higher accuracy. The size and shape of the predicted areas determine the model’s expressiveness (user gain) while influencing the degree to which the model provides a correct prediction (reliability). In this chapter we will introduce the area localization score (ALS) as a novel metric for measuring the trade-off between expressiveness and reliability. We will explore different models for indoor area localization ranging from adaptive classification of predefined segments, bounding box regression and polygon prediction by incorporating the building model within the learning phase and evaluate the approaches with the introduced ALS.