Deep Learning-based identification of human gait by radar micro-Doppler measurements

Abstract

The radar micro-Doppler (m-D) signature of human gait has already been used successfully for a few classification tasks of human gait, for instance walking versus running and determining the number of humans under observation. owever, the more challenging problem of personnel identification has not been solved yet. The aim of this study is to prove that the human walking gait differs between individuals and that it can be used for personnel identification using CW X-band radar measurements. This study investigates the effect of human walking gait characteristics such as speed and stride as well as the gender on leading to distinctive m-D signatures. Both simulated data and measurements of 22 subjects walking from and towards the radar were used. Unsupervised earning based on Adversarial Autoencoders was used to map the m-D ignatures to a latent space. T-Distributed Stochastic Neighbor Embedding and Uniform Manifold Approximation and Projection were then used for clustering and visualization. This study shows that even very slight changes in the walking gait characteristics mentioned above lead to distinctive m-D signatures mapped into closely located points in the latent space. A VGG-16 convolutional neural network was used to identify the walking subjects based on their easured m-D signature. Accuracy of above 93.5% was achieved, proving that CW X-band radar m-D signature of human walking gait can be used for accurate personnel identification which is reliable for 22 participants.