A Method for Implementing LSTM-Based Multiple-People Identification System for Non-Contact Health Monitoring on Small-Scale FPGA

We propose a method for implementing an LSTM-based multiple-people identification system on a small-scale FPGA. To reduce the hardware-resource usage, we introduce a digit adjustment and bit reduction method for converting the floating-point type to the fixed-point type with high precision. It enables not only to reduce the hardware-resource usage without degrading the identification accuracy but also to shorten the computation time by 50% compared to that of the conventional floating-point type.

I. Introduction

As the average life expectancy of human has increased all over the world, health monitoring becomes more important. Data of vital-signs (heartbeat, respiration, etc.) are essential for healthcare and usually obtained by medical equipments such as electrocardiographs or smart watches. However, the wearable type gives physical and mental stress to care-receivers, and a distance between a caregiver and care-receiver should be kept enough due to spreading COVID-19. Therefore, sensing vital signs without any contact is desirable for safe and trusted health monitoring.