ION GNSS 2012
Session F2: Urban & Indoor Alternatives: Wireless

Title: Real Time Location Using WiFi Fingerprints and Multiple Targets Tracking
Author(s): L. Rapoport, Z. Junping, Y. Youwen, Huawei Technologies, Russia
Date/Time: Wednesday, September 19, 2012, 2:35 p.m.
Room: Grand Ballroom Center (Renaissance)

GNSS is widely used for outdoor location and navigation. GPS was successfully augmented by GLONASS. In future this constellation will be enhanced by Galileo, Compass, QZSS. Large satellite constellation, improvement of ephemeris quality, and using triple frequency band pseudo-range and carrier phase measurements will potentially make robust standalone solution to converge to sub decimeter level of accuracy for most part of outdoor navigation applications. Assisted GNSS expands its capabilities to harsh conditions, but nevertheless GNSS is blind in deep indoor environments. So other approaches to positioning have been proposed for indoor navigation. One of most widely used approaches is based on using 802.11 Wi-Fi. It allows for obtaining of low cost location solution using already deployed access points (AP) just modifying firmware part of the system. Received signal strength indicator (RSSI) is widely used in navigation equations as raw measurements. Classic logarithmic loss function connecting distance to AP with RSSI readings can be used only in ideal environments, free of sources of fading and multipath. Understanding of this disadvantage gave rise to using fingerprint-based methods, connecting samples (or fingerprints) of RSSI measurements with certain calibration (or anchor) points whose precise position is stored in the calibration database together with RSSI fingerprint vectors. The whole navigation sheme consists of two stages. At the first stage the calibration data base is collected. At the second stage the navigation engine receives RSSI real time data and consults with calibration data base for appropriate fingerprint samples. In the present paper both stages are addressed: 1) Construction of the calibration database. The most challenging issue is density of the calibration grid, allowing for re-construction of the whole radio-map with necessary detail, yet not making it excessively cumbersome. Calibration grid is proposed to be not uniform, it is made taking into account obstacles, and it is controlled in real time using piece-wise linear approximation of the RSSI radio-map based on the Delaunay triangulation. Predicted location accuracy provides useful feedback helping to deploy calibration points correctly solving a tradeoff between the grid density and expected accuracy. 2) Real time location. Having the RSSS measurements vector, the list of triangle - candidates for possible location is first constructed. Least squared scheme is then used to estimate location - candidate in each triangle. If there are too many candidates, they are condensed into several clusters. Typical number of clusters is 2-5. Then a bank of Kalman filters is updated the same way as that is made in the multiple target tracking. Every solution candidate updates the appropriate filter. If certain filter is not updated for several time instances, its predicted covariance matrix trace exceeds certain threshold and this certain branch is cut and the filter is deleted from the filters bank. On the other hand, if a solution candidate can not update existing filter from the filter bank, it gives rise of the new branch. Parameters of the filter depend on the expected dynamics, like "static", "pedestrian", "wheeled robot".

The following questions are addressed in detail in the present paper: a) Delaunay triangulation of the calibrated area taking into account obstacles described by polygons taken from the floor plan, b) RSSI radio field approximation inside the triangle, comparison of different approximation schemes, c) Construction of solution candidates clusters, d) Construction of KF filters bank and multiple candidates tracking and updating scheme, e) Navigation engine architecture and multiple users tracking. f) Results of experiments are discussed and analysed for different testing scenarios, day time, and environments.

Previous Abstract Return to Session F2 Next Abstract