Radar-Based Efficient Gait Classification using Gaussian Prototypical Networks

Remote gait sensing and classification has potential applications in medical diagnosis, assisted living and recognition tasks. In this paper, we propose to classify human's gait data extracted from radar exploiting the time-frequency representation of the micro-Doppler signatures. Gait recognition for two tasks, namely human identification and human with luggage type, are demonstrated in this paper that uses the cadence-velocity diagram (CVD) of the micro-Doppler time data as input to Gaussian prototypical network for classification. Gaussian prototypical networks learn the projection of CVD gait data into embedding vector along with covariance representing the confidence region around the embedding vector, which are then clustered for classification. We demonstrate the performance of our proposed solution with 8 individuals (33 +/- 7 years) for the task of person identification and human walking types namely usual walking, walking with backpack and walking with trolley.