RF Micro-Doppler Classification with Multiple Spectrograms from Angular Subspace Projections

Radio Frequency (RF) sensors present distinct ad-vantages over cameras or wearables for hand gesture recognition providing high resolution radial range and velocity measurement, being able to operate in dark and through the objects with high temporal and frequency resolutions. Moreover, the flexibility of the complex formatted data allows users to develop their own algorithms to generate various data representations such as time-frequency (Micro-Doppler - μD) maps, or range-Doppler or - angle as a function of time. However, conventional μ-D generation does not regard the angular information of the multiple targets existing in the RF data. Hence, multiple targets with different μ-D signatures at various angular positions create a mixed spec-trogram output reducing recognition performance. This paper proposes an angular projection approach on radar data cubes (RDCs) to generate raw radar data for defined angular subspaces. Hence multiple μ-D spectrograms for each angular subspace can be constructed from the projected data. The proposed approach has been tested on RF data for gross body movement and American Sign Language (ASL) recognition. It has been showed that the utilization of angular projected spectrograms increases classification accuracy for ASL and achieves recognition accuracy of 92.6% for 20 word ASL signs.