Recently, positioning systems such as GPS [1] have become very popular and have made possible a large range of applications. However, most actual techniques, especially those requiring satellite coverage, are not suitable for indoor positioning. And, as nearly all modern buildings are equipped with Wi-Fi access points, indoor positioning using IEEE 802.11 standard has now become a realistic alternative. Moreover, recent smartphones are commonly equipped with Wi-Fi sensors, which makes them adequate devices to implement such an indoor positioning system. The range of potential applications is very large. Indoor positioning systems could be used to give access to an interactive map of a building. For example, they could orientate a person through an airport to the boarding gate, help a student find his classroom or facilitate the way of finding items of a shopping list in a supermarket. One successful approach for indoor positioning is based on Wi-fi fingerprints. It is applicable to scenarii with severe multipath unlike triangulation techniques where the distance to the base-stations need to be estimated based on time-of-arrival, roundtrip-time or signal strength attenuation [3]. Moreover, those techniques often require uninterfered propagation paths to work well. The fingerprint-based algorithms work differently and contain two phases: an offline and an online phase. The purpose of the offline phase is to collect information about the Wi-Fi access points signal strengths at different locations. During the online phase, the measured signal strengths are compared to the offline measurements in order to estimate the user position. For example, the positioning system RADAR [4] uses the Euclideandistancebetweenvectorsofstrengthsasasimilaritycriterion while the conditional joint probabilities are suggested in [5] and [6]. In an attempt to improve the accuracy of fingerprint-based indoor positioning systems, we propose a new method that compares online and offline signal strength probability distributions in order to find the nearest offline locations. Contrary to other techniques for which the signal strengths are averaged, we take advantage of the signal strength variations by considering the whole probability distributions. When applying the RADAR [4] and LOCATOR [6] methods to our testing data, we find that our method is about 1m more accurate at the 50% level of the CDF of the positioning error. This article is organized as follows: in Section 2, the principles of the method are described. Section 3 gives the main experimental results and Section 4 proposes a comparison with other techniques.