Low Power Radar-based Air-Writing System using Genetic Algorithm-assisted Spiking Legendre Memory Unit

In this paper, we propose a novel radar-based air-writing system using spiking neural network (SNN), enabling user to input characters into a user interface by drawing on imaginary board. In contrast to conventional deep learning (DL) approaches, SNNs are power-efficient and suitable for low-cost battery-operated IoT operations. We further propose Genetic Algorithm (GA) to find the optimal parameters of Spiking Legendre Memory Unit (SLMU) for the given application. Compared to their deep learning equivalent, the proposed solution offers similar level of accuracy 98.53% in case of two radars and outperforms in case of single radar with an accuracy of 95.37%, with estimated energy consumption of 2.04µJ. This is significantly less than the DL counterparts that can consumes energy in order of mJ. Additionally, the proposed systems require less storage memory 490kB for single radar and 564.2kB for two radars compared to the state-of-the-arts solutions, thus having smaller memory footprints.