Spiking Neural Networks for Gesture Recognition Using Time Domain Radar Data

Gesture recognition using luminance invariant radar sensors is vital due to its extensive use in human-machine interfaces. However, the necessity for computationally expensive radar data pre-processing steps represented by fast Fourier transforms to get range and Doppler features are regarded as a contemporary concern. In this work, we present a solution for gesture recognition that relies on time-domain radar data applied to an event-driven, sparse, and end-to-end trained spiking neural network architecture. Using the proposed solution, it is possible to discriminate between 10 different gestures in a gesture dataset recorded using a 60 GHz frequency-modulated continuous-wave radar sensor, with a mean test accuracy of 93.1%.