Cooperative and Distributed Localization for Wireless Sensor Networks in Multipath Environments

We consider the problem of sensor localization in a wireless network in a multipath environment, where time and angle of arrival information are available at each sensor. We propose a distributed algorithm based on belief propagation, which allows sensors to cooperatively self-localize with respect to one single anchor in a multihop network. The algorithm has low overhead and is scalable. Simulations show that although the network is loopy, the proposed algorithm converges, and achieves good localization accuracy.

💡 Research Summary

The paper addresses the problem of node localization in wireless sensor networks (WSNs) operating in multipath (non‑line‑of‑sight, NLOS) environments where each node can measure both time‑of‑arrival (TOA) and angle‑of‑arrival (AOA). Traditional GPS‑based solutions are unsuitable for indoor or power‑constrained sensor nodes, and most existing WSN localization methods assume line‑of‑sight (LOS) conditions. Moreover, many NLOS‑focused techniques treat the NLOS error as a random bias and rely on computationally intensive Bayesian inference (e.g., particle filters or bootstrap sampling).

The authors propose a novel distributed algorithm that combines a geometric model of multipath propagation with belief propagation (BP) on a factor graph. The geometric model considers R possible propagation paths (LOS or single‑bounce NLOS) between any pair of nodes Si and Sj. For each path r, the measured distance dr ij (derived from TOA) and the measured angles θr ij, θr ji (derived from AOA) are related to the unknown node positions through a linear equation of the form

 dr ji = g(θr ij, θr ji)ᵀ (si − sj)

where g(·) is a known vector function. By stacking the measurements from all R paths, the authors obtain a linear system

 dj i = Gj i si − sj + noise

with Gaussian noise of variance σ². The pseudo‑inverse G† ij of the measurement matrix Gij yields a closed‑form estimate of si given sj:

 si = sj + G† ij (dj i − noise).

If multiple neighboring nodes are available, the joint posterior distribution of all node positions becomes a high‑dimensional product of Gaussian likelihoods. Direct MAP estimation would require inverting a large matrix whose size grows with the number of nodes, which is impractical for large networks.

To overcome this, the authors construct a factor graph where each variable node represents a sensor position si and each factor node represents the pairwise likelihood between two neighboring sensors. The pairwise likelihoods are Gaussian because the measurement model is linear with Gaussian noise. Belief propagation is then applied: each node sends a Gaussian “message” (mean and covariance) to its neighbors, and updates its own belief by multiplying the incoming messages. The anchor node S0 has a known position (0,0) and therefore its belief is a Dirac delta; all other nodes start with a broad Gaussian prior (zero mean, large variance).

Because all messages remain Gaussian, only two scalars per dimension (mean vector and 2×2 covariance matrix) need to be exchanged, resulting in very low communication overhead. The update equations are derived in closed form (Equations 10‑14). The algorithm proceeds iteratively: at each iteration each node broadcasts its current belief, receives beliefs from its neighbors, computes new messages using the linear measurement model, and updates its belief. Convergence is typically achieved within a few tens of iterations even when the factor graph contains loops.

Simulation experiments are conducted with five sensors randomly placed in a 10 m × 10 m area. The anchor S0 is fixed at the origin. The ranging error is modeled as zero‑mean Gaussian with σ = 3 m, and the AOA error is uniformly distributed in