Proposal of polarimetric remote sensing systems based on quaternion self-organizing map for classification of human motions

This paper proposes polarimetric remote sensing systems based on quaternion neural network for classifying human motion. By focusing on polarization states, we can reduce the effects of fluctuations in electromagnetic wave intensity, and extract features by treating the polarization states as time-series signals. There, we represent the polarization states on or in the Poincare sphere in three dimensions via Stokes vectors. The systems perform unsupervised learning by employing a quaternion self-organizing map (QSOM), which is suitable for learning the relationship among three-dimensional information, and is particularly good at learning rotations in three dimensions, enabling flexible motion classification based on polarization time series. In experiments, we find that the number of required virtual-transmission polarization can be small as four, and that the time window size can be set flexibly, although there is an optimum value for the motion set that we want to detect. These results indicate sufficient technical feasibility of QSOM polarimetric systems for human monitoring.