Exploiting Sparsity of Ranging Biases for NLOS Mitigation

We study robust network localization for realistic mixed line-of-sight and non-line-of-sight (LOS/NLOS) scenarios, where (i) NLOS identification is not performed, (ii) no statistical knowledge of the LOS/NLOS measurement error is available, and (iii) no experimental campaign is affordable. We treat the bias term of each range measurement, both for LOS and NLOS, as an unknown parameter. Based on this, we indicate that the ranging biases possess a sparsity property in LOS-heavy scenarios. To exploit this sparsity, we propose the inclusion of a sparsity-promoting term into the conventional cost functions, giving rise to a generic sparsity-promoting regularized formulation. By bounding the cost function, we further develop an alternative generic bound-constrained regularized formulation. To ensure global optimality, we specify the residual error function in these formulations so that they are conveniently solved via relaxation as two semidefinite programs (SDPs). It is also shown that the two SDPs can be equivalent in the sense that they share the same optimal solution. Compared with the sparsity-promoting regularized SDP, the bound-constrained regularized SDP has the advantage that it allows us to develop one data-driven strategy for selecting an appropriate regularization parameter. Numerical results, based on both synthetic- and experimental data, demonstrate the overall enhanced performance of the devised approach, both in terms of localization accuracy and computational efficiency. The remarkable ability of the proposed data-driven method for parameter selection, at the cost of a slight increase in computational complexity, is also shown.