Fusion of Multiple Datasets to Support Location-based Services in Retail Applications

The accuracy of a positioning system is often coupled with its costs, e.g. a costly (and dense) infrastructure leads to a higher positioning accuracy. A current trend in indoor positioning are so-called opportunistic systems, which use an already existent infrastructure, e.g. WiFi access points. These systems are known for an accuracy in the range of several meters. In a retail scenario, this accuracy may be sufficient for so-called macro navigation (finding the right area in a shop of a specific item) but insufficient for micro navigation, i.e. finding an item in a shelf. Especially location-based services cannot be established due to the inaccuracy of the positioning. In addition to this, installed sensors in such environments can detect user interactions, however they often cannot identify the person who is interacting. For example, in our Innovative Retail Lab (IRL), shelves are instrumented with RFID readers, which enable the detection of product placing or removal without identifying the interacting user. In this paper, we describe a method that enables the fusion of such an opportunistic positioning system with shelf interactions and userrelated information, such as the contents of a shopping list, to derive a better position estimation on the one hand and an identification and disambiguation of user interactions on the other. This fusion approach is based on Dynamic Bayesian Networks (DBNs). The paper points out how the fusion can be used to provide location-based services.