A mmWave MIMO Radar-Based Gesture Recognition Using Fusion of Range, Velocity, and Angular Information

• Published in: IEEE sensors journal Vol. 24; no. 6; pp. 9124 - 9134
• Main Authors: Yu, Jih-Tsun, Tseng, Yen-Hsiang, Tseng, Po-Hsuan
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.03.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Angular resolution; Angular velocity; Artificial neural networks; Azimuth; Chirp; Continuous radiation; Convolutional neural networks; Deep learning; Feature extraction; Gesture recognition; human–computer interaction; Machine learning; millimeter wave (mmWave) radar; Millimeter wave communication; Millimeter waves; MIMO communication; multiple-input–multiple-output (MIMO) radar; Radar; Radar imaging; Sensors; Velocity
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3355395

Radar sensing technology offers an innovative approach to human-computer interaction, distinguished by its robust sensing capabilities impervious to acoustic and optical disturbances, thus presenting a superior alternative for reliable user engagement. Gesture recognition with millimeter wave (mmWave) frequency-modulated continuous wave (FMCW) radar extracts the range and velocity from the raw data, such as range-Doppler image (RDI). Besides, the angle estimated using a multiple-input-multiple-output (MIMO) radar also contains rich gesture information. Thus, to leverage the MIMO radar technique, we design the usage of the azimuth/elevation-based range-angle images (RAIs) with the average over slow time with the RDI as the spectrum map input for gesture recognition. Since gesture motion is related to the track of position and velocity, we utilize feature extraction on convolutional neural networks (CNNs) of two different spectrum maps, each time sequence learning by cascading long short-term memory (LSTM), and then fuse the two networks in the end to recognize hand gestures. We validate the proposed scheme based on hand gestures collected by several subjects in different rooms using the 77-GHz mmWave radar from Texas Instrument (TI). With the various combinations of antennas, we observed that the higher angular resolution provided by azimuth and elevation angles could help to enhance the machine-learning model's ability. By utilizing angle alongside velocity information, the late fusion network provides a classification accuracy of 94.67% and 97.43% among 12 gestures in terms of per frame and sequence accuracy.