Abstract
This paper presents a robust, distributed Expectation Maximization (EM) algorithm designed to localize non-cooperative radio frequency (RF) targets in challenging Non-Line-of-Sight (NLOS) environments. By modeling unknown signal reflections as latent variables and employing a gossip-based communication strategy, the proposed method achieves localization accuracy comparable to centralized systems while enhancing network resilience and energy efficiency.
Introduction
In real-world urban environments, obstacles frequently block the direct Line-of-Sight (LOS) between sensors and targets. Signals instead propagate via Non-Line-of-Sight (NLOS) paths, reflecting off surfaces (scatterers) before reaching the sensor network. Standard localization algorithms often fail in these scenarios due to the complex geometry of reflected paths.
Furthermore, traditional localization systems are typically centralized, requiring all sensors to transmit raw data to a single fusion center. This architecture suffers from high energy consumption at the hub and represents a single point of failure. This research addresses these issues by developing a distributed framework for localizing non-cooperative targets using Time-Difference-of-Arrival (TDOA) and Angle-of-Arrival (AOA) measurements.
Methodology
The proposed solution centers on a Distributed Generalized Expectation Maximization (EM) algorithm.
The EM Framework
The EM algorithm is employed to find maximum likelihood estimates in the presence of latent variables—in this case, the unknown orientations of the scatterers.
- E-Step (Expectation): Each sensor node calculates local summary statistics based on its current estimate of the target's location and the conditional probability of various scatterer angles.
- M-Step (Maximization): Nodes update their target position estimates to maximize the likelihood of the observed data.
Gossip-Based Consensus
To avoid centralized processing, the algorithm utilizes a Gossip Scheme. Following the local E-step, nodes exchange summary statistics only with immediate neighbors. Through iterative communication, the entire network converges toward a consensus on the target's location. A significant technical challenge addressed is the non-exponential nature of the data distribution, which the authors resolve using a sequential optimization technique in the M-step to ensure convergence to a local optimum.
Key Insights and Identifiability
A critical finding is that while TDOA measurements alone are often insufficient in NLOS conditions, the integration of AOA data significantly enhances localization accuracy. By treating reflection angles as latent variables, the algorithm can infer environmental geometry without a prior map.
However, the problem of "identifiability" arises where multiple potential target locations (ghost targets) may satisfy the measurements. To resolve this, the orientation of at least one reference node's scatterer must be known or estimated.
Experimental Evaluation
The approach was validated through both numerical simulations and an experimental campaign. The real-world setup utilized four sensor nodes based on the USRP-N210 software-defined radio platform. The test environment included one LOS node and three NLOS nodes, with signals reflecting off buildings or diffracting around corners. The system was evaluated across 13 distinct target locations.
Results and Discussion
Experimental results demonstrate that the distributed algorithm achieves a localization error of less than 15 meters for 90% of measurements, despite TDOA errors as large as 160 meters. Simulation data confirms that the distributed method's Root Mean Square Error (RMSE) closely matches that of a centralized system, proving that decentralization does not necessitate a significant loss in precision.
While the current model assumes single-bounce reflections, future work may extend this to multi-bounce paths in more complex indoor settings. Additionally, while the study focused on static targets, the iterative nature of the gossip protocol suggests potential for tracking mobile targets in real-time.
Summary updated