WO2016079656A1

WO2016079656A1 - Zero-calibration accurate rf-based localization system for realistic environments

Info

Publication number: WO2016079656A1
Application number: PCT/IB2015/058850
Authority: WO
Inventors: Moustafa Amin Youssef; Rizanne MAMDOUH
Original assignee: Egypt-Japan University Of Science And Technology
Priority date: 2014-11-18
Filing date: 2015-11-16
Publication date: 2016-05-26

Abstract

A system that can provide zero-calibration accurate RF-based indoor localization that works in realistic environments on heterogeneous devices. It leverages a novel incremental Voronoi tessellation approach that reduces the user ambiguity region by applying successive RSS constraints and can achieve a median distance error of 2.5m compared to traditional fingerprinting technique. This system is robust to changes in the environment and devices and can provide at least 33% percent enhancement in accuracy over prior art systems based on outdated fingerprints.

Description

Zero-Calibration Accurate RF-based Localization

System for Realistic Environments

Field of the Invention

[0001] This invention relates to the field of location-based services and more specifically to radio frequency (RF) based location determination systems applicable to indoor use.

Background of the Invention

[0002] The widespread use of wireless networks combined with ubiquitous

mobile devices has led to the proliferation of RF-based localization techniques, which are able to provide a solution for indoor tracking using received signal strength (RSS) readings from a plurality of local RF transmitters, such as access points (APs). Some examples of such systems include WiFi-based localization and the more recent iBeacons based localization that leverages Bluetooth Low Energy (BLE) technology,

[0003] Typically, RF-based indoor localization systems require a tedious

calibration phase of the area of interest, where a RF fingerprint is constructed to capture the noisy and complex indoor propagation of RF signals. This fingerprint requires continuous updates to capture temporal and spatial changes in the environment and dynamic changes of the transmit power of the APs to handle interference or dynamic load change on the AP. Moreover, the heterogeneity of user devices introduces a further challenge of capturing the hardware differences between these different devices in the fingerprint.

[0004] To address these challenges, a number of RF localization techniques have been proposed that attempt to reduce the calibration overhead through crowd-sourcing of the fingerprint construction that require explicit or implicit user feedback, using CAD tools or propagation models to automate the fingerprint construction process, or combining RF localization with other sensors. These systems, however, have a trade-off between accuracy and overhead, still require some form of calibration, and usually incur higher energy consumption and/or cost.

[0005] To handle the heterogeneity of devices, a number of approaches that map the fingerprint constructed by one device to another have been introduced. Nevertheless, the range of available user devices in the market, which grows each day, leads to inaccurate mapping in some cases and requires some training and/or a learning process for each new device.

[0006] Finally, to address the dynamic power changes of smart APs (i.e., APs that can dynamically change their power outputs) and other temporal or spatial changes, special sniffers have been used to monitor the area of interest, increasing the deployment cost of the localization system. These challenges highlight the need for a system that is calibration- free, accurate, robust to heterogeneity in user devices, and adapts to dynamic changes in the environment and the transmit power of APs.

Summary of the Invention

To address the shortcomings of the current state of the art, a zero- calibration accurate RF-based indoor localization system that works in real environments on heterogeneous devices is presented.

Referring to Figure 1 , the system utilizes the fact that the relation between the RSS and distance from the transmitter is monotonic. That is, the further away the receiver is from the transmitter, the lower the signal strength and vice versa.

Given this, a receiver at a particular location hearing multiple APs must, generally, be closer to the AP from which the strongest signal is received than to the other APs. If a Voronoi Tessellation is constructed using the APs as seeds, this is generality the equivalent of placing the receiver in the Voronoi cell containing the strongest AP.

Therefore, the system starts by constructing the Voronoi diagram of the area of interest, taking the APs as the seeds, as shown in Figure 2(a). To further reduce the ambiguity of the user location, spatial constraints are applied on the user location based on the pairwise RSS relation of every pair of APs. Specifically, given the RSS from two APs, RSS(_4) and RSS(5), if RSS(/i) is greater than RSS(£?), then the receiver has to be closer to AP(/i) than AP(B) and vice versa, as shown in Figure 1. In 2D space, this maps to the half plane where AP(/i) is located. Given that m APs can be heard at the current receiver location, we can have up to l ) pairwise constraints.

Some of these constraints define the boundary of the initial Voronoi cell in which the receiver is located and some of them are redundant given other constraints. Applying all of the pairwise constraints incrementally further reduces the ambiguity region of the receiver. The final location of the receiver is estimated as the center of mass of the final ambiguity region, as shown in in Figure 2(d).

Calculating the constraints in realtime may be too computationally intensive to run on battery constrained devices (receivers), such as mobile phones. To address this issue, a gridding approach is used, where a virtual grid is super-imposed on the area of interest, as shown in Figure 3. For each grid point, the expected constraint value for each pair of APs is calculated offline based on the physical distances between the grid point and the different APs. While the system is running in realtime, all grid points are scanned and the actual constraint value based on the RSS from the two APs is compared to the expected constraint value stored for each grid point. The estimated receiver location is taken as the center of mass of the grid points that have the largest number of matching expected and actual constraints. [0014] The density of the grid points can be chosen given the tradeoff between computational complexity and accuracy. In addition, due to noisy wireless channels that can lead to fluctuations of the RSS values, in some cases some constraints may be conflicting, leading to a null feasible receiver region in the case of the original incremental Voronoi Tessellation

algorithm. The virtual gridding algorithm solves this problem by

estimating the receiver location based on the grid points that match the largest number of actual and expected constraints.

[0015] A system using the incremental Voronoi Tessellation with the virtual

gridding extension presented herein allows a calibration free solution for indoor localization to be obtained. In addition, the system can handle heterogeneous user devices as it only depends on the relative RSS values between pairs of APs, rather than the absolute signal strength.

[0016] The system also includes a number of modules to handle practical deployment issues including the noisy wireless environment, obstacles in the environment that may affect the relative RSS relations, heterogeneous device hardware, smart access points with dynamic power control, 3D positioning of APs (i.e., APs positioned at different heights), accuracy estimation, and efficiently computes the constraints in real-time.

Brief Description of the Figures

[0017] Figure 1 shows that for any pair of APs, if the RSS from AP(A) at the

device is greater than the RSS from AP(B), then the device must be closer to AP(A) than to AP(B). The device then can be mapped to the half plane defined by the bisector line between AP(A) and AP(B) and containing AP(A).

Figures 2(a-d) shows the basic approach using four APs.

Figure 3 shows virtual gridding approach extension of the present invention

Figure 4 shows that a wall between the AP(A) and the receiver leads to reducing the RSS from AP(A), as compared to the case where no walls exists between the receiver and the AP(A).

Figure 5 shows the architecture of the system of the present invention.

Figure 6 shows the handling of the effect of walls on the received signal strength.

Without loss of generality, a 2D area of interest is assumed where multiple APs have been installed. A user carrying a device at an unknown location 1 scans for nearby APs. The scanned APs' information can be represented as a vector s = [si,...,s_m], where m is the number of heard APs and each si is an ordered pair of (APi, RSSi) representing the ID and signal strength of the i^th heard AP. The entries in the vector s are sorted in a descending order according to the RSS. To reduce the computational overhead, the system uses a virtual gridding approach. Let gi represent the i^th2D grid point. TheVoronoi cell associated with the i^thAP is denoted as Vori.

Referring to Figure 5, the system works in two phases: an offline phase and an online tracking phase:

Offline phase:

During the offline phase, the system administrator uses the User Interface module to enter a floorplan of the area of interest tagged with the locations of the APs and the walls. The information regarding wall locations is optional, and can be obtained easily from the building CAD information. The inclusion of the wall location information helps in reducing the impact of the building structure on the system accuracy.

The virtual grid is generated in this phase using the Grid Generator module. The density of the grid can be configured by the user to trade off accuracy and computation overhead. In the preferred embodiment of the invention, the grid points are evenly spaced, but in other embodiments, this constraint is unnecessary.

The P re-Computation module calculates the parameters associated with each grid point. Preferably, the associated parameters are the expected pairwise constraints evaluation based on the distance between the grid point and each pair of APs. This helps in reducing the running time during the online tracking phase. The Pre-Computation module also calculates the Voronoi diagram of the area of interest. This is used for calculating the initial user ambiguity area as well as clustering the grid points during the online phase to reduce the computation overhead.

The initial ambiguity area can be calculated using the basic Voronoi Tessellation, as shown in Figs. 2(a-d). Figure 2(a) shows the initial step where the ambiguity region of the location of the receiver is the Voronoi cell corresponding to the AP with the highest RSS.

Figure 2(b) shows applying the first constraint, that being the RSS at the receiver from AP(B) and AP(C). Because RSS(B) > RSS(C), the receiver must be closer to AP(B) than to AP(C). This implies that the receiver falls in the half plane defined by the bisector line between AP(B) and AP(C) and which contains AP(B).

Likewise, Figure 2(c) shows applying second constraint, that being the RSS at the receiver from AP(A) and AP(B). The receiver falls in the half plane defined by the bisector line between AP(A) and AP(B) and containing AP(A).

This gives the ambiguity region shown in Figure 2(d). The estimation of the final location of the receiver can be provided as the center of mass of this ambiguity region.

Online tracking phase: This is the main operational phase of this system, where receivers can be tracked in realtime. The process starts by scanning for the APs that can be heard at the current device location.

The Pre-Processor compensates for the dynamic power control of the different APs and smooths the input signal to reduce the wireless channel noise. It also selects a subset of the grid points to evaluate, further reducing the computational overhead. The Constraint Evaluator then compares the expected RSS constraint with the stored distance constraints for each pair of heard APs for each grid point. The output of this module is a list of each grid point and its associated number of matched constraints. The Location Estimator fuses the output of the Constraint Evaluator module to obtain a single user location.

Finally, the Accuracy Estimator associates a confidence with the estimated location reflecting the accuracy of the estimated location.

The system contains the following main modules:

Pre-Processor. The pre-processor has three goals: a. compensating for the dynamic power control of the different APs, b. handling the noise in the wireless channel, and c. selecting a subset of the grid points to evaluate in the rest of the modules. The power normalizer compensates for different power outputs of the APs. Different APs can be configured to transmit at different powers to achieve different coverage and interference goals. This difference in transmit power can be static or dynamic based on the environment condition as in smart APs and can lead to violating the uniformity assumption between the different APs. To address this issue the system leverages the information available from the different RF technologies.

Specifically, smart APs usually give access to their configuration information which can provide the current transmit power of the different APs. Similarly, the iBeacons technology includes the transmit power in the transmitted frame and this information is available from the APIs on both the Android and iOS operating systems. The present invention leverages this information to remove the transmit power offset by subtracting it from the scanned RSS vector.

The noise handler compensates for noisy wireless channels. Due to the noisy wireless channel and the complex signal propagation in indoor environments, the RSS can fluctuate with time, even at a fixed location, which may lead to swapping the constraints of two APs, especially when their RSSs are close. To help reduce this effect, the pre-processor analyzes a window of RSS samples, rather than a single RSS value. Different filters can be used including the average and median filters. [0043] The probabilistic constraint evaluator gives accuracy from 5-9% enhancement in median error as it takes the full RSS histogram into account when evaluating the constraints compared to the median and average techniques that do not use the full available information.

[0044] To reduce the computational requirements of the system, the pre-processor also selects all or a subset of the grid points for constraints evaluation. The idea is that the final user ambiguity region (and hence the most probable grid points) will be a subset of the grid points within the Voronoi cell with the strongest heard AP. In the preferred embodiment of the invention, all grid points within a pre-determined number of the strongest Voronoi cells are evaluated. The pre-processors select the points inside these Voronoi cells as the initial grid points for evaluating the RSS pairwise constraints. To further address the noise effect of APs with close RSS, the Voronoi cells associated with the strongest APs are used as the initial grid points.

[0045] The handling of obstacles in the environment is shown in Figure 6. For a certain grid point and two specific APs, different number of walls can exist between this grid point and the two APs. Because each wall leads to attenuation of the signal that goes through it, this leads to violating the uniformity assumption between the different APs. To address this issue, the present invention estimates the number of walls, Walls i ), between each grid point i and the AP j in its vicinity using the Pre-Computation module during the offline phase. During the online tracking phase, the Constraint Evaluator module uses the well-known wall attenuation factor model to compensate for the number of walls between an AP and a grid point. In particular, it transforms the input RSS scanning vector s by

adding a constant factor for each wall between the AP and the grid point.

More formally, for grid point i and AP_j, the transformed RSS value

RSS(/ is obtained from the input RSS RSS(/) as:

RSS(/ = RSS(/^') + Walls( ) C (1 ) where C is a constant parameter representing the signal attenuation due to a single wall.

This transformation is applied to all APs heard at a specific grid point, unifying the effect of the different number of walls between them, as shown in Figure 6.

The Constraint Evaluator module compares the expected distance

pairwise constraints calculated during the offline phase with the actual

RSS constraints in realtime for each selected grid point from the preprocessing module. It outputs, for each grid point, the number of matched expected and actual constraints. The present invention also introduces two novel contributions to further enhance the accuracy and handle the noisy RSS, Non-heard APs constraints and probabilistic constraints evaluation.

Non-heard APs constraints:

To further enhance accuracy, it is noted that a non-heard AP provides further information about the location of the user. Specifically, each non-heard AP adds new set of m constraints, one with each one of the heard APs. That is, the user must be closer to a heard AP than a non-heard AP. These additional constraints further reduce the user ambiguity region. To reduce the computational overhead introduced by the additional constraints the present invention only leverages the non-heard AP within a fixed radius from the strongest AP.

[0052] Probabilistic constraints evaluation:

[0053] Due to the noisy non-deterministic nature of wireless, it is better to use a

probabilistic approach for comparing the RSS of each pair of APs. Instead of relying on the average or median RSS only, the system of the present invention calculates the probability that the RSS received from one of the APs within a window of samples is higher than the other. This is based on the RSS histograms within a time window w. More formally, for two APs A and B, Pr(RSS_A > RSS_B) can be calculated as:

[0054] If Pr(RSS_A > RSS_B) is greater than 0.5, AP(A) is considered to be the

strongest AP. Otherwise, AP(B) is considered to have a higher RSS.

[0055] The User Location Estimator module aims to estimate the final user

location from the candidate weighted grid points (weighted by the number of matched pairwise constraints) by using three techniques: Maximum matching, k-highest, and p-highest. Maximum matching technique: The maximum matching technique estimates the user location as the center of mass of the grid points that have the maximum number of matching constraints, regardless of their number. In a perfect environment, these should correspond to the grid points within the final user ambiguity area. k-highest technique: This technique estimates the user location as the

center of mass of the top k% of locations, weighted by their matching

score, where k is a fixed parameter. p-highest technique: Instead of using a fixed number of grid points like the previous technique, this technique estimates the user location as the center of mass of the top locations whose sum of probabilities (i.e. weights)

exceeds a certain parameter p.

To further enhance accuracy, the User Location Estimator averages the last t estimates to obtain a smoothed output.

The accuracy estimator, as shown ion Figure 5, associates a confidence level with the estimated location reflecting the accuracy of the estimated location, using one or more of several methods.

Variance-based Method: Under the assumption that the localization error is a zero mean Gaussian and the user is not moving, the error variance is equal to the variance of the estimated location. The circle radius can be estimated as the one (two) sigma value, which should make the estimated location within the given circle 68% (95%) of the time. [0062] Grid-based Method: In this method, the radius of the circle is estimated proportionally to the distance between the estimated user location and the furthest grid point that were used in the center of mass calculation to estimate this location.

[0063] As shown in Figure 5, the invention can be performed directly on the mobile device (i.e., the receiver) and the results sent to the entity doing the tracking. A typical such mobile device would be a smart phone, equipped with a processor capable of running an application implementing the invention, and an RF transceiver for connecting to networks from which the received signal strength from a plurality of RF transmitters may be obtained. Alternatively, the computations may be performed on a server, with the necessary information being regarding received signal strength being sent to the server from the mobile device. The information generated in the offline phase is typically stored on a server and accessible by the mobile device, but in some embodiments may be kept in local memory on the mobile device.

[0064] The invention is also not limited to APs using Wi-Fi. The scope of the invention is meant to be broad enough to cover any type of RF signal, including, but not limited to, Wi-Fi, Bluetooth and cellular. Note also that the term "AP" or "access point" as used throughout is meant to refer to any RF transmitter and is not meant to be limited to RF transmitters that can be used to connect to a network.

Claims

We claim:

1. A system for localizing a mobile device having an RF receiver in an indoor setting, comprising: a. a plurality of RF transmitters disposed at various locations within said indoor setting; b. a virtual grid overlaid on said indoor setting, said virtual grid consisting of an array of grid points corresponding to physical locations in said indoor setting and, associated with each grid point, a plurality of parameters representing the expected distance pairwise constraints evaluation based on the distance between the grid point and each pair of RF transmitters; and c. a processor, configured with software comprising: i. a pre-processor module for performing the functions of compensating for the dynamic power output of the RF transmitters, handling the noise in the wireless channel, and selecting a subset of said grid points to evaluate; ii. a constraint evaluator module for comparing said expected distance pairwise constraints with the actual received signal strength constraints for each pair of RF transmitters for each of said selected grid points and outputting, for each of said selected grid points, a weighting based on the number of matched expected and actual constraints; iii. a location estimator module for estimating the location of said mobile device from said selected weighted grid points; and iv. an accuracy estimation module, for associating a confidence level with the estimated location.

2. The system of claim 1 wherein said software further comprises: v. a user interface module, said user interface module accepting from a user: a floorplan of said indoor setting; and the location of said plurality of RF transmitters within said indoor setting.

3. The system of claim 2 wherein said user interface further allows the user to enter the density of said virtual grid.

4. The system of claim 1 wherein said array of grid points is evenly spaced over said indoor setting.

5. The system of claim 1 wherein said software further comprises: vi. a pre-computation module for calculating: for each of said plurality of grid points, said plurality of parameters representing the expected distance pairwise constraints evaluation based on the distance between the grid point and each pair of RF transmitters; and a Voronoi diagram of said indoor setting, using each of said RF transmitters as seeds.

6. The system of claim 5 wherein said pre- computation module further performs the function of calculating an initial ambiguity area.

7. The system of claim 1 wherein said pre-processor module further comprises: a power normalizer for performing the functions of: obtaining a transmitting power reading from all or a subset of said RF transmitters; and normalizing said transmit power readings to compensate for different power outputs of said RF transmitters.

8. The system of claim 1 wherein said pre-processor module further comprises: a noise handler for performing the functions of: analyzing a window of the received signal strength readings from all or a subset of said RF transmitters; determining any reduction in said received signal strength due to and compensating for said reduction in received signal strength.

9. The system of claim 8 wherein said determining function applies an average or median filter across said window of received signal strength readings.

10. The system of claim 8 wherein said determining function applies a probabilistic constraint evaluator over a full received signal strength histogram.

11. The system of claim 1 wherein said pre-processor module selects all or a subset of said grid points for evaluation by selecting all or a subset of said grid points within a predetermined number of Voronoi cells having the strongest received signal strength.

12. The system of claim 5 wherein said pre-computation module further performs the function of estimating the number of walls between each of said grid points and each of said RF transmitters.

13. The system of claim 6 wherein said constraints evaluator module uses information about RF transmitters from which no signal is being received to refine said initial ambiguity area.

14. The system of claim 1 wherein said location of said mobile device is determined by estimating said mobile device as the center of mass of a subset of said grid points having the maximum number of matching constraints, said subset of grid points defining final user ambiguity area.

15. The system of claim 1 wherein said location of said mobile device is determined by estimating said mobile device as the center of mass of the top k% of said grid points weighted by a matching score.

16. The system of claim 1 wherein said location of said mobile device is determined by estimating said mobile device as the center of mass of the top grid points whose weights exceed a predetermined parameter.

17. The system of claim 1 wherein said accuracy estimator uses a variance-based method wherein the error variance is equal to the variance of said estimated location and further wherein said location of said mobile device is within a circle having a radius defined a sigma value of a Gaussian function.

18. The system of claim 1 wherein said accuracy estimator uses a grid-based method wherein said location of said mobile device is within a circle having a radius estimated proportionally to a distance between the estimated location of said mobile device and the furthest of said grid points used in a center of mass calculation.

19. A system for localizing a mobile device having an RF receiver in an indoor setting, comprising: a. a plurality of RF transmitters disposed at various locations within said indoor setting; and b. a processor and memory, associated with said mobile device, said processor configured with software for performing the functions of: i. constructing a Voronoi diagram of said indoor setting, using said plurality of RF transmitters as seeds; ii. creating an ambiguity region as a cell in said Voronoi diagram having the strongest received signal strength; iii. refining said ambiguity region by calculating spatial constraints based on the received signal strength from each pair of RF

transmitters; and iv. estimating the location of said mobile device as the center of mass of said refined ambiguity region.

20. The system of claim 19 further comprising: a virtual grid overlaid on said indoor setting, said virtual grid consisting of an array of grid points corresponding to physical locations in said indoor setting; wherein an expected spatial constraint value is calculated for each pair of RF transmitters with respect to each of said grid points; wherein said mobile device scans all or a subset of said of said grid points and compares an actual constraint value based on received signal strength from one or more pairs of said RF transmitters to said expected constraint values for each of said scanned grid points; and wherein the location of said mobile device is estimated as being the center of mass of a predetermined number of grid points having the largest number of actual and expected constraint values.

21. The system of claim 20 wherein said expected spatial constraint values for each of said grid points are pre-calculated and stored in said memory of said mobile device.