Noncontact Multi-Finger Micro-Gesture Recognition Using Frequency-Modulated Continuous-Wave Ultrasonics and Piezoelectric Micromachined Ultrasonic Transducer Arrays

Noncontact ultrasonic sensing using piezoelectric micromachined ultrasonic transducers (PMUTs) is promising for compact, low-power, and robust human–machine interaction (HMI). However, accurate recognition of multi-finger micro-gestures remains a great challenge because of PMUTs’ weak echoes, limited bandwidth, and multi-target interference. To tackle these issues, we develop a noncontact multi-finger micro-gesture recognition system using frequency-modulated continuous-wave (FMCW) ultrasonics and PMUT arrays in a one-transmit–two-receive configuration. Range–Doppler maps (RDMs) are generated from FMCW echoes to jointly characterize target range and velocity, while a phase-difference map (PDM) is constructed to encode spatial orientation information from dual-receiver measurements. To learn the resulting spatiotemporal representations, an attention-enhanced network termed Atten-3DCNN-LSTM is developed by integrating collaborative spatiotemporal attention, three-dimensional convolutional neural networks (3DCNNs), and long short-term memory (LSTM) units. Experiments on seven representative multi-finger micro-gestures achieve an average user-dependent random-split accuracy of 96.43%, outperforming the conventional 3DCNN and 3DCNN–LSTM baselines, while leave-one-subject-out (LOSO) cross-subject evaluation further achieved an average accuracy of 81.93%. Our research demonstrates effectiveness of the proposed system for high-precision ultrasonic micro-gesture recognition in HMI applications.