Previous Abstract Return to Session C6 Next Abstract

Session C6: Terrestrial Signals of Opportunity-Based Navigation Systems

Heteroscedastic Gaussian Process Model for Received Signal Strength Based Device-Free Localization
Ossi Kaltiokallio, Unit of Electrical Engineering, Tampere University; Roland Hostettler, Department of Electrical Engineering, Uppsala University; Jukka Talvitie, and Mikko Valkama, Unit of Electrical Engineering, Tampere University

Radio frequency (RF) sensing technologies utilize perturbations of the wireless radio channel for estimating physical quantities of interest such as presence of people, crowd density, activity, gestures, location and vital signs. The attributes “device-free”, “passive” and “sensorless” are typically used to highlight that the technology does not require the monitored subject, which we refer to as target from now on, to carry or wear any active or passive device. The technology leverages the fact that targets alter the propagation characteristics of radio signals, and these changes can be quantified at the receiver (RX) using the radio channel estimate of the radio module. Then, the physical quantity of interest can be inferred from the channel estimate(s) using for example a parametric statistical model or a non-parametric machine learning model. Over the past two decades, various wireless technologies have been used to demonstrate the capabilities of RF sensing. As an example, dedicated hardware combined with large bandwidths can be used for acquiring high resolution delay estimates enabling for example remote breathing and heart rate monitoring. At the other end of the spectrum are commodity wireless devices that provide received signal strength (RSS) estimates. Even though the RSS is not as informative as delay or phase estimates, it still conveys useful information that can be used to realize various RF sensing applications such as breathing monitoring and localization. The main benefit of using inexpensive commodity wireless devices is that they can be deployed in numbers forming a dense mesh network and to perform multistatic sensing. In this paper, we consider a multistatic sensing system, composed of commodity narrowband wireless devices capable of measuring the RSS, for device-free localization and tracking (DFLT).
The performance of state-of-the-art DFLT systems is significantly influenced by the measurement model that describes the RSS as a function of target’s location. Significant research efforts have been undertaken to model the RSS using, for example, first principles, and it has been argued that the target induced perturbations to the wireless channel are caused by shadowing, reflection, and/or diffraction. In addition, a wide variety of empirical models have been proposed. Common to all these models is that the largest RSS changes are measured in between the transmitter (TX) and RX, and that the influence decays as the bistatic range to the target increases. A major limitation of the analytical and empirical models is that they are unable to explain RSS changes when the target alters an existing multipath component or when the bistatic range is large. Another significant drawback is that the models typically assume homoscedastic measurement noise, that is, the RSS is assumed to be corrupted by independent and identically distributed (i.i.d.) random noise. However, both theoretical and empirical evidence suggest that the noise process is heteroscedastic, that is, the measurement noise is dependent on the target’s location. Fingerprint-based machine learning approaches can capture the RSS changes as well as the heteroscedastic noise appropriately. However, collecting the fingerprints is laborious and the models are non-continuous allowing estimation only at discrete fingerprint locations.
In this paper, we propose a novel heteroscedastic GP model for RSS-based DFLT. GPs provide a Bayesian, non-parametric approach for data driven modeling of smooth functions. The main benefit of GPs is that they allow us to approximate highly nonlinear signal propagation models, uncertainty can be correctly handled, the model is continuous allowing estimation at arbitrary locations, and prior knowledge and beliefs can be fused into the GP model. The main contributions of the proposed GP model include: i) we relax the constraint of i.i.d. random noise and model the heteroscedastic noise using a GP; ii) we explicitly model the mean function which allows us to integrate our prior knowledge on the propagation mechanisms into the GP model which also improves interpretability of the model; and iii) we incorporate prior beliefs on hyperparameters by deriving the maximum a posteriori (MAP) estimate for the hyperparameters. The proposed model expresses that the RSS is close to a global linear model with the residuals as well as the noise being modeled by GPs. To the best of our knowledge, this is the first work that bridges the gap between parametric statistical models and non-parametric machine learning models since it combines properties from both approaches.
In this work, we also implement two different tracking filters to solve the DFLT problem. The first filter is a two-step approach. In the first step, the target is first localized using a maximum likelihood estimate (MLE) which is computed over all TX-RX pairs of the multistatic sensing system. Then in the second step, the MLE is used as input to a Kalman filter to track the kinematic state of the target. The second filter is a nonlinear particle filter which can directly use the RSS as well as the proposed GP model to track the target.
The development efforts of the paper are demonstrated using commodity radios that operate on the 2.4 GHz ISM band and according to the IEEE 802.15.4 standard. However, it is important to note that the developed methods can be generalized to any device capable of measuring the signal strength including Wi-Fi, Bluetooth and RFID. The experiments are conducted in an open indoor environment and in a downtown residential apartment and in both experiments, 20 sensors are deployed into the environment. Using the experimental data, it is shown that the proposed approach can decrease the localization error up to 77% with respect to a benchmark parametric model and up to 51% with respect to a benchmark non-parametric model. The experimental results imply that the localization root mean squared error (RMSE) is 14.2 cm in the open indoor environment and 29.4 cm in the challenging residential apartment experiment when using the proposed model. Furthermore, the RMSE decreases notably when filtering is applied to track the kinematic state of the target. Thus, the experimental results validate the research premise that GPs can capture complex propagation patterns that cannot be explained with analytic models as well as target dependent measurement noise – resulting in superior performance with respect to state-of-the-art parametric and non-parametric models.

Previous Abstract Return to Session C6 Next Abstract