Gaussian mixture models for signal mapping and positioning
Introduction
Indoor positioning is often based on signals of existing wireless networks due to the lack of GNSS signals, high cost of installing and upkeeping dedicated positioning infrastructure, and wide availability of the wireless networks [1]. However, the use of existing wireless infrastructure that is not designed for indoor positioning, such as WLAN or Bluetooth, requires a map of the signal environment [2], a so-called “radio map”. Also, consumer WLAN systems provide only RSS measurements and, for example, range [3], time-of-arrival [4] or direction-of arrival [5] measurements, cannot be used for positioning.
Because walls attenuate and reflect electromagnetic signals and antenna radiation patterns are not directionally uniform, signal strengths cannot be modeled accurately with omnidirectional models, such as path loss in vacuum. Signal attenuation can be modeled using analytic or numerical physics models that take into account the location and materials of the walls [6]. An alternative to physics based models is an empirical model based on a set of measurements of RSS values collected at known locations in the indoor region of interest. These measurements are commonly called FP, and there exist many fingerprint positioning methods [7], [8], [9], [10], [11].
In this paper, we concentrate on situations where each FP is associated with the location where the measurement was made and the radio map is built first in an offline phase. The map is later used in the online phase for positioning. In the literature, there are also algorithms that map the environment without knowing FP locations [12], [13], [14], [15], doing the mapping and positioning in one phase.
One approach to estimate the radio map from FP is WKNN. Here the position is computed by comparing RSS values with FP in the database; the estimated position is a weighted average of locations of the k FP with the most similar RSS values among received AP [16]. The learning phase, thus, consists of collecting the FP. In the positioning phase, the received RSS values are compared to FP in the database. In this phase, there are several variations of weighting the k nearest points. The weighting may be based on the difference of RSS values [16] or the rank of the RSS [17]. The data may also be preprocessed to improve the positioning performance in a varying signal environment [18]. WKNN is commonly used due to its simplicity, but it has two relevant shortcomings: its accuracy can be improved, it does not provide error boundaries for its estimates, and the database size and computational cost increases as more FP are collected.
Kernel methods and histogram methods [16] extend the WKNN by modeling the distribution of RSS values at calibration points. In the learning phase, multiple measurements are made at each mapped location and the distribution is modeled using a histogram or a kernel method. This increases the required work in the mapping phase significantly and gives only small improvement in the database size or accuracy.
In [19], missing measurement values in FP are filled and then principal component analysis is applied for dimensionality reduction. Then a GMM is fitted to model the reduced dimension FP. Compared to WKNN the computational complexity of the positioning does not increase as the number of samples increase.
In [20], a GMM is proposed to be used to model a the distribution of RSS values inside cells. Compared to the kernel and historgam methods this method allows to track the dependency of RSS values from different AP. In [21] a GMM is used to model the RSS values in the mapped area. The GMM is not used as probabilistic model, but rather as a measurement model with an added Gaussian noise.
PL models are based on a functional relation between signal strength and the distance or path to the receiver [22], [23]. Isotropy may be assumed in open space with omnidirectional antennas. But in built environments, the walls and other structures attenuate and reflect the signals. In particular, the attenuation caused by concrete structures is larger than attenuation from light indoor walls [24]. This anisotropy causes problems if a PL model is to be used in multifloor buildings.
In proximity method the measurement is hard thresholded to provide information whether the measurement was close to an AP [25]. Similarly coverage area methods consider only area where an AP is received with strong enough RSS [26].
GP were used to model radio maps in [27], [28], [29]. Their approach has the advantage that it can model RSS patterns of arbitrary distribution in space. However, they need to assume Gaussian noise and the variance of the noise does not depend on the measured RSS values.
Among the above presented methods, only kernel, histogram and GMM methods in [19], [20] are able to model non-Gaussian variations of RSS values, which is useful for quantifying measurement uncertainty in a probabilistic positioning algorithm. However, these methods can only model this characteristic at the points or cells where multiple measurements where collected.
In this paper, we propose a model that fills the above-described gap in the literature. The model uses a GM for mapping RSS values in buildings and the map is used for indoor positioning. Specifically we propose a GMM for the joint distribution of RSS values and locations. The main benefit, compared to other methods in literature, is that it models the distribution of the RSS values for each AP in the mapped area without the need of making multiple measurements at each FP location. The proposed model allows us to compute location estimates and RSS distributions in closed form, which allows to evaluate models without approximations.
Compared to GMM based algorithm in [20], the proposed algorithm has the benefit that there is no need for collecting samples in cells as the position dimension is considered as a continuum. Similarly [19] treats the measurement points as separate points without using spatial information. Furthermore, the algorithm in [21] uses the GMM only as a measurement model.
Compared to the Gaussian process methods in [27], [28], [29], the proposed model and GP models can model arbitrary signal attenuation patterns inside buildings, but the proposed model can also model arbitrary distributions of RSS values measured at any location close to the area where measurements were made, while GP models rely on the Gaussian distribution.
One important aspect of the model is that it can be used to estimate the distribution of the attenuation and of RSS values in different parts of the building. This is valuable information also for other applications than indoor positioning; for example, RSS maps can help radio infrastructure planners to identify areas that have poor data transfer rates.
The rest of this paper is organized as follows. Section 2 gives the theoretical background of our approach. Section 3 presents the proposed model for building an RSS map. The filtering algorithm based on the proposed model is presented in Section 4. Section 5 presents examples of applying the proposed for WLAN signal mapping and positioning in a four story building. Section 6 concludes the paper.
Section snippets
Problem formulation
A Gaussian mixture is a pdf of formwhere k is the number of components, pN(x|μi, Pi) is the pdf of a multivariate normal distribution with mean μi and covariance Pi and weights wi are positive and sum to 1. A Gaussian mixture can approximate any pdf as accurately as desired by increasing the number of components [30].
GMM can be used in various ways for position estimation. They can be used in Bayesian filters [30]. The generation of components can be automated based on
Modeling RSS maps with GMs
In this section, we explain how to model RSS maps with GMs. In Section 3.1, we introduce the RSS model. In Section 3.2, we indicate how we can find the parameters for the GMM.
Filtering using GMs
In this section, we present the GMF that will be used for the positioning. In Section 4.1, we show the formulas for computing conditional distributions from GMM for the GMF and the actual GMF is presented in Section 4.2.
Experimental results
In order to test the proposed algorithm we use FP that were measured from 4 floors of a campus building at Tampere University. Data was collected by a person who walks around the buildings and manually enters the true locations when collecting RSS data. Data was collected with a tablet computer and the number of fingerprints in each floor is: 1530 (Floor 1), 1583 (Floor 2), 333 (Floor 3), and 107 (Floor 4).
Conclusions and future work
The main novelty this work is a new GMM for WLAN RSS mapping. The new model models the 3D position and RSS values in a single augmented multidimensional distribution. The use of a GMM allows to compute the conditional pdf in closed form. The main findings of the article are:
- •
Proposed GMM can model the anisotropic attenuation and varying distribution of RSS values inside a building without the requirement to make multitude of measurements at each point.
- •
In real world positioning tests, the
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
Financial support by the Academy of Finland under the grant no. #295080 (CrowdSLAM) is hereby gratefully acknowledged.
References (45)
- et al.
Generalized rejection sampling schemes and applications in signal processing
Signal Process.
(2010) - et al.
A multi-model sequential Monte Carlo methodology for indoor tracking: Algorithms and experimental results
Signal Process.
(2012) - et al.
A survey of indoor positioning systems for wireless personal networks
IEEE Commun. Surv. Tutor.
(2009) - et al.
WLAN location determination via clustering and probability distributions
Proceedings of the First IEEE International Conference on Pervasive Computing and Communications, 2003. (PerCom 2003).
(2003) - et al.
Relative location estimation in wireless sensor networks
IEEE Trans. Signal Process.
(2003) Angle of arrival estimation using WiFi and smartphones
Proceedings of the International Conference on Indoor Positioning and Indoor Navigation (IPIN)
(2016)- et al.
Propagation modeling for accurate indoor WLAN RSS-based localization
2010 IEEE 72nd Vehicular Technology Conference - Fall
(2010) - et al.
A survey of parametric fingerprint-positioning methods
Gyrosc. Navig.
(2016) - et al.
Indoor tracking: theory, methods, and technologies
IEEE Trans. Veh. Technol.
(2015) - et al.
A survey on indoor positioning systems
2014 22nd International Conference on Software, Telecommunications and Computer Networks (SoftCOM)
(2014)
Evolution of indoor positioning technologies: a survey
J. Sensors
A survey of selected indoor positioning methods for smartphones
IEEE Commun. Surv. Tutor.
Efficient, generalized indoor WiFi GraphSLAM
2011 IEEE International Conference on Robotics and Automation
Indoor localization without a prior map by trajectory learning from crowdsourced measurements
IEEE Trans. Instrum.Meas.
Evaluation of two WiFi positioning systems based on autonomous crowdsourcing of handheld devices for indoor navigation
IEEE Trans. Mobile Comput.
WiFi-SLAM using Gaussian process latent variable models
A comparative survey of WLAN location fingerprinting methods
Proceedings of the 6th Workshop on Positioning, Navigation and Communication 2009 (WPNC’09)
Rank based fingerprinting algorithm for indoor positioning
International Conference on Indoor Positioning and Indoor Navigation (IPIN), 21–23 September 2011, Guimarăes, Portugal
Combining similarity functions and majority rules for multi-building, multi-floor, WiFi positioning
2012 International Conference on Indoor Positioning and Indoor Navigation (IPIN)
CEnsLoc: infrastructure-less indoor localization methodology using GMM clustering-based classification ensembles
Mobile Inf. Syst.
High-dimensional probabilistic fingerprinting in wireless sensor networks based on a multivariate Gaussian mixture model
Sensors
Design and implementation of WiFi indoor localization based on Gaussian mixture model and particle filter
2012 International Conference on Indoor Positioning and Indoor Navigation (IPIN)
Cited by (0)
- 1
Member of EURASIP