Robust subspace detectors based on weighted least-squares

In this paper, we propose and design robust subspace detectors for classification of multidimensional subspace signals. Using the principle of M-estimators and least-median-of-squares (LMedS), we formulate the robust subspace detectors as weighted subspace detectors, where we weigh the rows of the measurement matrix prior to the signal matching. The detectors are demonstrated numerically by communication signals transmitted over an unknown frequency selective channel in impulsive noise, and shape classification of partially occluded two-dimensional objects. In both cases, the proposed robust subspace detectors outperform the classical subspace detector.