Mmblink:Anti-accompanying Interference Blink Detection System Using Deep Millimeter Wave Radar Sensing

Blink detection plays a crucial role in application scenarios such as eye disease detection,human-computer interaction,fatigue prevention driving,and emotion perception.To address the unique characteristics of micro-scale,sparse and non-periodic blinking,especially the failure of millimeter wave radar-based blink detection caused by the accompanying interference(breathing,heartbeat,small-scale head motion)from the human body mixed with a blink in a non-linear manner.This paper proposes a selfsupervised deep contrastive learning method based on a non-linear independent component analysis framework to separate the accompanying interference from the blink signal.The research in this paper is summarized as follows:(1)The study explores the feasibility of millimeter wave radar in blink detection and establishes a mathematical model for detecting blink displacement by measuring the phase of the radar IF signal.Also,the interaction between human companion interference signals and blink signals is investigated;(2)This paper proposes a solution based on a non-linear independent component analysis framework for the problem of nonlinearly mixed accompanying interference and blink signals that traditional methods cannot solve.The solution designs a time-correlated,single-feature separation network ES-Net1,which successfully achieves the separation of nonlinearly mixed signals by taking two positive and negative sample sequences with time correlation and time uncorrelated as the input of the network and using the feature extractor inside ES-Net1 to recover the temporal structure of blink and accompanying interference signals;(3)To further improve the accuracy of signal separation,this paper designs a timedependent double-feature separation network,ES-Net2,and adds one feature time series as the network input.With the increase of input features,the dataset construction of the double-feature separation network becomes more diverse.To adapt to the new dataset structure,a two-channel network is used to extract the features of both feature sequences to improve the measurement accuracy of the system;(4)This paper implements the proposed prototype system for blink detection against accompanying interference based on TI’s IWR1642 millimeter wave radar platform.Experimental results show that the signal separation capability of ES-Net1 and ES-Net2 is up to 85.75% and 90.4%.The average blink frequency error rates under different accompanying interference are 7.9% and 8.3%,respectively.The comprehensive experimental results verify that the system can achieve stable blink detection in the presence of human accompanying interference.