The real-time position and mobility of a user are the most crucial information for personalized location-based services (LBSs) in the era of the Internet of Things (IoT). With the pervasiveness of mobile devices (MDs), people can enjoy satisfactory LBSs in outdoor environments in general by using the embedded Global Positioning System (GPS) module on MDs. However, the location estimates of GPS modules are inaccurate in densely populated and highly built-up urban areas.
Realizing equal levels of LBSs in indoor environments is still a challenging task since the performance of GPS modules degrade tremendously due to the lack of light of sight propagation channels. Various wireless sensing modalities, e.g., RFID, Bluetooth, and WiFi, have been proposed for indoor localization. With the pervasiveness and wide availability of WiFi infrastructure, and nearly every MD being equipped with a WiFi module, WiFi has been acknowledged as the most promising alternative to GPS for indoor LBSs.
The most promising algorithm for a WiFi-based indoor positioning system (IPS) is a fingerprinting-based algorithm. This consists of 2 steps: 1) in the offline training phase, receiving signal strength (RSS) measurements from multiple WiFi access points (APs) at numbers of reference calibration points (CPs) and collecting their coordinates and forming them as fingerprints to establish a radio and spatial mapping database over the indoor environment; and 2) in the online implementation phase, estimating the location of a MD by searching and matching the real-time RSS readings to the fingerprints stored in the database.
A key reason that a fingerprinting-based algorithm provides higher localization accuracy than a signal propagation model-based approach is that it captures the irregular signal variations in complex indoor environments. The quality of the offline radio map is vital for the localization performance of fingerprinting-based algorithms Thus, manual RSS data collection and site survey approaches are commonly adopted. However, constructing the radio map is extremely labor-intensive and time-consuming. In addition, the manually calibrated radio map at one time is vulnerable to temporal and environmental dynamics. These are two major bottlenecks that hinder WiFi-based IPS for ubiquitous indoor LBS implementation.
Various methods have been proposed to tackle these issues in recent years. For instance, researchers propose to automatically collect RSS measurements in a crowdsourcing manner when users are performing daily activities. A major drawback of crowdsourcing is the location uncertainties of the RSS data, which may lead to inaccurate spatial and radio mapping. Therefore, the localization accuracy of crowdsourcing method is usually lower than that of a fingerprinting-based algorithm with a manually calibrated radio map.
Another approach is to apply machine learning methods, e.g., manifold alignment and Gaussian process regression (GPR), to train a model with RSS data on a limited number of calibration points and generate interpolated RSS values on other locations for radio map construction. However, the scalability of GPR is poor and its representational power is restricted by the limited choice of kernel functions, e.g., squared exponential kernel. These functions are too smooth to mimic signal propagations in complex indoor environments that are severely affected by multipath effects caused by furniture, walls, and moving occupants.
Generative adversarial networks (GANs) can achieve huge successes in generative applications, especially in synthesizing realistic images in recent years. Thus, a tensor-GAN (TGAN) can be proposed to use GAN to generate high resolution representation of real RSS measurements at calibration points. However, this does not estimate RSS values at new coordinates and, therefore, does not enhance the spatial granularity of the radio map.
The disclosed technology is generally directed to an automatic radio map construction and adaptation scheme to model the irregular Global Positioning System (GPS) signal distribution in these area for better location estimations. The automatic scheme may construct a fine-grained radio map over the entire area of interest. Implementations of the disclosed technology can include an automatic wireless fine-grained ratio map construction and adaptation scheme that is empowered by the designed Gaussian process regression (GPR) initialized Wasserstein generative adversarial network (GAN). In certain embodiments, any environment can be divided into two categories: a free space where a signal measurement sensor can access freely, and a constrained space the sensor cannot access or where the noise level of its measurement is high.
The disclosed system, referred to herein as WiGAN, may firstly establish a GPR model using the real RSS readings collected by the sensors in multiple calibration points in the free space. After that, the coarse RSS estimations from the GPR can be used as inputs for the generator of the GAN, and the generator and discriminator of the GAN can be trained in a minimum-maximum fashion. The discriminator may aim to identify whether the RSS data is real or synthesized, while the generator may construct realistic RSS data to fool the generator. In this manner, the irregular RSS distributions in complex environments can be modeled by inheriting the advantages of both the GPR and the GAN. To establish a radio map in constrained space without any real data collection procedure, the trained GPR model can be used to generate coarse RSS estimations at those virtual points within the constrained space. Then, by using them as the input for the GAN's generator, the fine-grained RSS values at points inside the area of interest can be estimated from the generator.
Implementations of the disclosed technology, referred to herein as WiGAN, are generally directed to an automatic fine-grained ratio map construction scheme that is empowered by a Gaussian process regression (GPR) initialized Wasserstein generative adversarial network (GAN). Any environment can be divided into two categories: free space where a signal measurement sensor can access freely, and constrained space where the sensor cannot access or the noise level of its measurement is high. The WiGAN may firstly establish a GPR model using the real RSS readings collected by the sensor in the free space. Then, the outputs of the GPR may be adopted as the input of GAN's generator.
A training objective of the GAN may be to model the irregular RSS distributions in the environment, and generate realistic RSS data in the constrained space that have not been manually site surveyed. The trained GPR model may be used to generate coarse RSS estimations at those virtual points within the constrained space. Then, by using them as the input for the GAN's generator, the fine-grained RSS values at points inside the area of interest may be estimated from the generator. The WiGAN can be leveraged to generate radio map for any suitable RF signals, e.g., GPS, LTE, BLE, and WiFi.
In certain implementations, WiGAN can be leveraged to generate a radio map for any RF signals (e.g., GPS, LTE, BLE). WiGAN may be used to construct a fine-grained WiFi radio map automatically in a complex indoor environment. The RSS distribution can be modeled with Gaussian Process Regression (GPR) and fundamentals of Generative Adversarial Network (GAN). After that, WiGAN can construct the fine-grained radio map by way of the GPR initialized Wasserstein generative adversarial network.
Gaussian Process Regression for Coarse RSS Estimation
GPR has been used in many applications, e.g., spatial smoothing, geostatistics, and robotics. Since its nonlinear regression capacity is more suitable to capture the anomalous RSS variations in complex indoor environments than linear signal delay function, researchers have leveraged it to interpolate RSS values and, consequently, reduce the cost for offline calibration.
Consider a mobile robot that collects RSS values from p access points (APs) at n calibration points in the free space. The spatial map and radio map dataset includes n pairs of (liƒ, Siƒ)i=1n where liƒ=(xi, yi) is the two-dimensional coordinates of a calibration points, and Siƒ is the RSS values from p APs at location liƒ. The relationship between the spatial space L and the signal space S can be modeled as:
S=ƒ(l)+{grave over (σ)}
where {grave over (σ)} is independent and identically distributed (i.i.d.) additive zero-mean Gaussian noise with variance σ{grave over (σ)}2. The latent function ƒ(l) can be modeled as a Gaussian Process (GP):
S˜GP(m(l),k(l,l′))
where m(⋅) and k(⋅,⋅) represent the mean and covariance function of GP respectively. For RSS radio map construction, the mean function m(⋅) can be modeled as a Log-Distance path loss model and a quadratic polynomial function. The quadratic polynomial function can be utilized as the mean function:
m(li)=β0+β1xi+β2yi+β3xi2+β4yi2+β5xiyi,
where (xi, yi) are the coordinates of location li, and the squared exponential kernel function as the covariance function:
where σƒ2, r and δ(⋅,⋅) denote the signal variance, the scale parameter, and the Kronecker delta function, respectively.
The Gaussian process regression model (GP) can be used to provide a coarse estimation of RSS values {Ŝjs}j=1m∈ŜGPs from p APs at m locations {ljs}j=1m∈Ls locations in constrained space. At each location ljs, the RSS value from each AP can be estimated according to the posterior mean of GP:
s
j
s
=m(ljs)+K(ljs,Lƒ)[K(Lƒ,Lƒ)+σ{grave over (σ)}I]−1(Sƒ−m(Lƒ)).
Generative Adversarial Network (GAN)
A GAN as described herein generally includes two parts: a generative model G and a discriminative model D. By taking a noisy vector z˜Pn(z) (e.g., Gaussian or uniformly distributed) as the input, the objective of the generator G is to synthesize data ({circumflex over (x)}) resembling real data distribution, x˜Pr(x). The discriminator D can randomly sample a batch of real and synthesized data and distinguish them by maximizing the probabilities that x is real and {circumflex over (x)} is fake. The loss function of GAN can be formulated as:
minGmaxDEx˜P
where G and D can be trained jointly on the loss function in a minimum-maximum fashion using backpropagation, which updates D to maximize the likelihood of the discriminator being correct while updating G to generate realistic data to fool the discriminator. One potential issue with the GAN is the vanishing gradients for the generator. Thus, WiGAN can be used to minimize the Wasserstein distance between the distributions instead of the original Jensen-Shannon divergence in the settings of GAN.
The GAN generally obtains groundbreaking results for the generation of realistic images. A few studies apply GAN on RF signal generations, e.g., audio signal and EEG signals. A Tensor Generative Adversarial Network (TGAN) can be used to generate a representation of RSS values with super-resolution for indoor localization.
WiGAN
WiGAN may generate realistic and fine-grained RSS values in constrained space (e.g., on a table of cubicles, conference rooms, and personal offices) using its generator G with the output of the GPR model GP learned using the data collected in free space (e.g., corridors, open space). The potential issue is more challenging since RSS values are to be estimated at new coordinates in order to construct a fine-grained radio map in the entire environment, while conventional methods focus on generating more RSS samples at the coordinates where real RSS measurements are already available.
An example of using WiGAN for radio map construction and adaptation is presented in Algorithm 1 and illustrated by
Step 1: a GPR model GP is constructed using the RSS values of p APs collected at n calibration points (CPs) in the free space as well as their 2D coordinates (Sƒ,lƒ).
Step 2: this step of WiGAN aims to train its generator G and discriminator D with (Sƒ,ŜGPƒ). In WiGAN, instead of using a noisy vector z˜Pn(z) as the input for G in GAN, the system concatenates the coarse RSS estimation from GP together with the noisy vector as ŜGPƒ and uses them as the inputs for G. In this manner, WiGAN inherits the nonlinear relations between the spatial and radio space captured by GP and has better initialization for adversarial training than using the random noisy vector. The inputs for the discriminator D are randomly sampled batches of real RSS data Sƒ˜Pr(Sƒ) and synthesized RSS data ŜGPƒ˜Pn(ŜGPƒ) in the free space. The loss function of WiGAN may be formulated as:
minGmaxDES
Backprogration can be used to optimize the parameters of G and D in the min-max loss function. A primary objective of G may be to generate realistic RSS data to fool D, while D may aim to maximize the likelihood of being correct of the determination of whether the data is real or fake. The Wasserstein distance can be adopted as the metric to measure the discrepancy between the two data distributions to improve the training stability of WiGAN.
Step 3: to estimate the RSS values at virtual points (VPs) in the constrained space, a coarse estimation ŜGPƒ can be provided via the GP model established in Step 1. Suppose there are m testing point in the constrained space, the coarse RSS value of each AP at each testing point (j=1, . . . , m) may be calculated as:
ŝ
GP
s
=m(ljs)+K(ljs,Lƒ)[K(Lƒ,Lƒ)+σ{grave over (σ)}2I]−1(Sƒ−m(Lƒ)).
Step 4: the coarse RSS estimations obtained ŜGPƒ by Step 3 together with a small random noisy vector may be utilized as the input for the generator G of WiGAN. The fine-grained RSS estimations at VPs in the constrained space may be calculated using the generator G trained in Step 2 as:
Ŝ
s
=G(ŜGPƒ).
In this manner, a fine-grained radio map that covers both free and constrained space may be established by WiGAN with higher spatial granularity.
Algorithm 1 below further describes the radio map construction and adaptation.
= m(ljs) + K(ljs, Lf)[K(Lf, Lf) + σs2I]−1(Sf − m(Lf))
Experimental Setup
To evaluate the radio map construction and adaptation performance of WiGAN, extensive experiments were conducted in a 700 m2 multi-functional office. As depicted in
An MD (here, an iPhone 7) was placed on a mobile robot (here, Turtlebot 2) to collect the RSS values of the MD automatically in the free space (e.g., corridors, open space) at 62 calibration points (highlighted as blue dots in
Fifty RSS samples were collected in each calibration point. The coordinates of the calibration points and the mean RSS value collected on each point (li, Si)i=1ƒ were utilized to construct the Gaussian process regression model. Then, the synthesized RSS samples and the real RSS values collected on the calibration points were used to train the generator and discriminator of WiGAN in the min-max manner as illustrated in
Evaluation on RSS Estimation Accuracy
The mean (ēRSS) and standard deviation (σRSS) of RSS estimation errors of WiGAN were calculated with respect to the observed RSS measurements at the 80 testing points in the constrained space. The mean RSS estimation error of WiGAN (80×50 samples) was 3.56 dBm with the standard deviation of 2.7 dBm on average. Its detailed performance at each AP was compared with GPR. As demonstrated in Table 1 (ēRSS) and Table 2 (σRSS), WiGAN reduced the mean error by 54.5% and the standard deviation of RSS error by 49.5% compared to GPR on average, which achieves tremendous performance gain. Statistical measurements of the two methods, including boxplot and histogram of RSS estimation error distribution of the 9 APs are illustrated by
Evaluation on Localization Accuracy
Since an objective of constructing and updating a fine-grained WiFi radio map is to improve the localization performance, the localization accuracy of the IPS was evaluated when WiGAN was utilized. Thirty testing points were randomly selected in the entire office. STI-WKNN was adopted as the localization algorithm together with different radio maps to estimate locations at these testing points and the localization errors (Euclidean distance) were calculated with respect to the physical locations (ground truth). The performance of WiGAN was compared with three radio maps, real RSS data collected by robot at the calibration points in free space (lower baseline), real RSS data collected by robot+generated RSS values via GPR, real RSS data collected by robot+generated RSS values via WiGAN, and real RSS data collected at calibration points in constrained space and free space (manually collected radio map).
Table 3 above compares the localization errors in terms of mean (ėLA) and standard deviation (σLA) of the four methods. As presented in Table 3, the use of RSS data collected by the robot in free space only is the worst (3.87 m) and cannot meet the localization accuracy requirement for many LBS applications. Both WiGAN and GPR enhance the accuracy over the lower baseline, which validates the necessity of generating RSS data in constrained space. A noteworthy observation from Table 3 is that WiGAN reduces the localization error by 33.5% over GPR, which indicates higher RSS estimation accuracy leading to higher localization accuracy. Another noteworthy point is the mean error of WiGAN reaching 1.98 m, which is only 0.2% lower than the manually collected radio map (upper-baseline).
The disclosed aspects may be implemented, in some cases, in hardware, firmware, software, or any combination thereof. The disclosed aspects may also be implemented as instructions carried by or stored on one or more or non-transitory computer-readable media, which may be read and executed by one or more processors. Such instructions may be referred to as a computer program product. Computer-readable media, as discussed herein, means any media that can be accessed by a computing device. By way of example, and not limitation, computer-readable media may comprise computer storage media and communication media.
Additionally, this written description makes reference to particular features. It is to be understood that the disclosure in this specification includes all possible combinations of those particular features. For example, where a particular feature is disclosed in the context of a particular aspect, that feature can also be used, to the extent possible, in the context of other aspects.
Also, when reference is made in this application to a method having two or more defined steps or operations, the defined steps or operations can be carried out in any order or simultaneously, unless the context excludes those possibilities.
Furthermore, the term “comprises” and its grammatical equivalents are used in this disclosure to mean that other components, features, steps, processes, operations, etc. are optionally present. For example, an article “comprising” or “which comprises” components A, B, and C can contain only components A, B, and C, or it can contain components A, B, and C along with one or more other components.
Also, directions such as “right” and “left” are used for convenience and in reference to the diagrams provided in figures. But the disclosed subject matter may have a number of orientations in actual use or in different implementations. Thus, a feature that is vertical, horizontal, to the right, or to the left in the figures may not have that same orientation or direction in all implementations.
Having described and illustrated the principles of the invention with reference to illustrated embodiments, it will be recognized that the illustrated embodiments may be modified in arrangement and detail without departing from such principles, and may be combined in any desired manner And although the foregoing discussion has focused on particular embodiments, other configurations are contemplated.
In particular, even though expressions such as “according to an embodiment of the invention” or the like are used herein, these phrases are meant to generally reference embodiment possibilities, and are not intended to limit the invention to particular embodiment configurations. As used herein, these terms may reference the same or different embodiments that are combinable into other embodiments.
Although specific embodiments of the invention have been illustrated and described for purposes of illustration, it will be understood that various modifications may be made without departing from the spirit and scope of the invention. Accordingly, the invention should not be limited except as by the appended.
This application claims priority to and the benefit of U.S. Provisional Patent Application No. 62/779,956, filed Dec. 14, 2018, which is incorporated herein by reference in its entirety.
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/US2019/066627 | 12/16/2019 | WO | 00 |
Number | Date | Country | |
---|---|---|---|
62779956 | Dec 2018 | US |