Human gait extraction from short and sparse radar dwells

This paper describes a novel approach to estimate two fundamental gait frequencies from a dismount observed by a radar that collects short dwells widely separated in time while it maintains wide area surveillance. For each dwell, range-Doppler images or chips are generated, and range versus range-rate maps are formed from few coherent processing intervals (CPIs) to obtain a spectrogram snapshot. The time duration of each spectrogram snapshot is about a quarter of the gait frequency cycle induced by the side to side movement of the body of the walking dismount, i.e., 0.25 s. At an instant of time, two spectrogram snapshots are fed to the algorithm; first, a limb mitigation filter is carried out on each spectrogram snapshot to extract the torso motion from which a Hankel matrix is computed. Next, the two Hankel matrices are concatenated into an augmented Hankel matrix to compute a set of state space matrices that give rise to the parameters associated with the dynamics of the dismount and the two fundamental gait frequencies. As time evolves, more spectrogram snapshots are generated and the gait frequencies are estimated to identify the observed dismount. Effectiveness of the gait extraction technique from sparse data is confirmed using simulated radar data from a six foot tall man that is generated using the Thalmann human motion model at six different aspect angles.